Professional Documents
Culture Documents
edited by
Xavier Lagrange
ENST Bretagne, Site de Rennes
and
Bijan Jabbari
George Mason University
A New Approach for Partitioning the Received SNR Space for Tractable 3
Performance Analysis in Wireless Packet Networks
M. Hassan, M. Krunz, W. Ryan
Packet Service in UMTS: Effects of the Radio Interface Parameters on the 103
Performance of the Downlink Shared Channel
F. Borgonovo, A. Capone, M. Cesana, L. Fratta
Cellular Multihop Networks and the Impact of Routing on the SNIR and 115
Total Power Consumption
K. M. Pepe, B. R. Vojcic
Approximate and Exact ML Detectors for CDMA and MIMO Systems: a 183
Tree Detection Approach
S. Vaton, T. Chonavel, S. Saoudi,
Block Turbo Code with Binary Input for Improving Quality of Service 195
P. Adde, R. Pyndiah, S. Kerouedan
Bounding Techniques for the Design of Channel Coding and Modulation 221
Systems
Y. W. Blankenship, B. K. Classon
Quality of Service of Internet Applications over the UMTS Radio Interface 239
S. Heier, A. Kemper, S. Grabner, lO. Rock
Third generation networks have been specified and are now being deployed
in a few countries. They are expected to reach maturity in the next several
years and to provide various services including audio, video, and world wide
web browsing. Furthermore, radio terminals are expected to be integrated in
a number of devices such as personal computers, personal digital assistants,
and even television sets. Such a wide-usage of radio mandates ongoing
research to address design of networks with high capacity while providing
acceptable quality of service.
This volume is the sixth in the edited series Multiaccess, Mobility and
Teletraffic for Wireless Communications. It presents the selected papers for
the proceedings of the Seventh Workshop (MMT'2002) held on this topic in
June 2002 in Rennes, France. The aim of this workshop has been to address
a set of important issues of interest to the wireless communications
community. In particular, the focus of this workshop is to identify, present
and discuss the theoretical and implementation issues critical to the design of
land based mobile cellular and microcellular as well as wireless local area
networks. Included in this book are recent research results on performance
analysis of wireless packet networks, channel coding and receiver design,
radio resource management in third generation systems, mobility
management in cellular and mobile IP networks, performance of transport
protocols (TCP) over radio link control protocols, and ad-hoc networks.
Xavier Lagrange
Bijan Jabbari
ix
MMT
Vol 6
X. Lagrange
B. Jabbari
A NEW APPROACH FOR PARTITIONING
THE RECEIVED SNR SPACE FOR TRACTABLE
PERFORMANCE ANALYSIS IN WIRELESS
PACKET NETWORKS
1. Introduction
Wireless networks are characterized by time-varying channels in which
the bit error rate (BER) varies dramatically according to the received
3
X. Lagrance and B. labbari (eds.),
Multiaccess, Mobility and Teletraffic for Wireless Communications, Volume 6, 3-24.
© 2002 Kluwer Academic Publishers.
4
,\'f(JteN
.l'll1teN-1
.fIate J
.flareO
~.
time
onloffsource
[]JJ]]rmm
Incoming trafic
N-l
2u rN-llu No
PCi = t
j=r+l
(r:) (Pe(i\))j (1 -
J
Pe(ri))n- j (1)
where n is the number of bits in a code block, including the FEC bits,
Pe(r) is the BER when the instantaneous SNR is r (the form of Pe(r)
depends in the underlying modulation scheme), and f\ is the "nominal"
SNR value in state i (its calculation will be discussed later). The packet
transmission/retransmission process can be approximated by a Bernoulli
process [1]. We assume that the transmitter always gets the feedback
message from the receiver before the next transmission slot, and a packet
is retransmitted persistently until it is successfully received. The nominal
service rate in state i, denoted by Ci, can be approximated by the inverse
of the mean of the geometrically distributed retransmission process:
Ci = c.e.(l - Pc;}, i = 0,1,2, ... , N (2)
where C is the error-free service rate, e = kin is the FEC overhead, and
k is the number of information bits in a code block.
Wireless transmission of continuous waveforms in obstacle environ-
ments is prone to multi-path fading, which results in randomly varying
envelope for the received signal. This randomness has been shown to fol-
Iowa Rayleigh distribution. In the presence of additive Gaussian noise,
the instantaneous received SNR r is proportional to the square of the sig-
nal envelope [5]. Accordingly, the SNR is exponentially distributed with
pdf:
1 -r
PR(r)=-e p , r>O (3)
p
(4)
8
7ri = l ri
ri +1 1 =
-e p dT
P
=e
::!:i
P - e
-ritl
P (5)
O(T")
, ~ cC((ooTi)) = Ni , i = 0,1, ... , N - 1 (6)
9
3. SNR Partitioning
It is easy to see that in the producer-consumer model, the steady-state
probability distribution is binomial, i.e.,
(7)
For our wireless channel to fit the Markov chain in [14], the partitioning
must be done so that no more than one state falls in the "good" (BER
close to zero) and "bad" (BER close to one) regions of the SNR space,
since either scenario will lead to an unrealistic number of states.
To completely specify the underlying Markov chain, we need to deter-
mine N, ~, and the thresholds Tl, T2, ... , TN. Our procedure for comput-
ing these quantities is outlined as follows. First, by equating (5) and (7),
we get an expression for N in terms of Tl and ~. Using the level-crossing
analysis and the structure of the embedded Markov chain, we then ob-
tain an expression for ~ in terms of Tl and TN. Combining the obtained
expressions, we can eipress N in terms of Tl and TN only. Then, by
selecting an appropriate value for TN, the value of the threshold Tl can
be obtained. After obtaining Tl, the other thresholds can be obtained
recursively, as follows: The steady-state probability that the channel is in
state i is given by (5), from which and after obtaining the ith threshold
Ti, the (i + l)th threshold can be obtained using the following expression:
The details of the above procedure are now presented. Let {X (t) : t 2
O} be the (irreducible) Markov process that represents the state of the
channel. The time spent in any state Si is exponentially distributed with
mean Ti = t. The parameter qi represents the total rate out of state Si·
Consider states 0 and N in Figure 3. The total rate out of state 0 can be
approximated by the level-crossing rate (LCR) at Tl, i.e.,
(9)
10
Similarly, the total rate out of state N can be approximated by the LCR
at rN:
(10)
(11)
(14)
From (5), and noting that rN+1 ~ 00, 1I'N can also be written as
:::!:.J:i..
1I'N = e p (16)
From (15) and (16), an expression for the ratio ~ in terms of rI, rN, and
N can be obtained: J.L
>. e:::!:.J:i..
p ) 11
( (17)
~= 1-e7
11
From (15), (16), and (11), we can write the following equation for 7fN:
(18)
(19)
where N is given by
N = -~-'---""':""":- (20)
In (1 + V7f :~ )
In our analysis, we choose r N such that the service rate at state N is
almost equal to the error-free service rate (this depends on the relation-
ship between the SNR and the BER, which in turn is dependent on the
modulation scheme). This is done by solving for rN in
Note that in step 9 of the algorithm if ()(Ti) > i/N for some state
i E {i, 2, ... , N -i}, then there is no Ti E [Ti' Ti+l) for which ()(ri) = i/N,
and the partitioning does not fulfill the requirements of the producer-
consumer model. If that happens, we decrement the value of Nand
repeat the computations (intuitively, decrementing N increases the ranges
of the various states, which improves the likelihood of finding appropriate
nominal SNR values). The recursion is step 8 is obtained from 7ri =
-r- -r-+l
e --t - e ~. Note that the algorithm is guaranteed to return a solution,
13
since for N = 1 the two nominal SNR values, 1'0 and 1'1, are given (the
partitioning reduces to the two-state Gilbert-Elliot model).
4. Performance Analysis
4.1 Queueing Model
Once the parameters of the FSMC are obtained, we proceed to com-
pute the queueing performance. The underlying queueing model can be
described by 2(N + I)-state Markov chain, as shown in Figure 5. The state
space of this chain is given by S = {(i, j) : i = 0,1 and j = 0,1,2, ... , N},
where a state (i, j) indicates that the source is in state i (on or off) and
the channel is in state j. Recall that the transition rates for the FSMC
are chosen such that the analysis in [13, 14] stays applicable.
"',I , ~,2
, )..i-I,i , , ,
.
N-2.N-I IN_I,N
n,n 0,1 O,H
.. O,i O,N-l
.. ',N
, • "'.I ,
... J.IN'-1,N-2 JlN,N-I
, , ,
~,o J.li,i-l
a a a a a p a
1,0
"',I ~
1,1
~,2
,
!.i-I
,-l,i ~
I.i
AN-2,N-l • l.N-1 "'-I,N ~
1,:-1
~ ~,o , 112,1
, I\,i-I
... )1N-I,N-2 ~ J.1N,N-l
Figure 5. State transition diagram of the Markov chain that governs the queueing
model.
where ai's are constant coefficients, and the pairs (Zi' ¢i), i = 1,2,3, ... ,
are the eigenvalues and left eigenvectors of the matrix MD- 1 (see [13, 14]
for details).
(26)
where AA(Z) and Ac(z) are the Gartner-Ellis limits for the accumulative
arrival process {A(t) : t ~ O} and the accumulative service process {C(t) :
t ~ O}, respectively, defined as
and similarly for Ac. Once z* is obtained, the packet loss performance
can be approximately given by:
where ACi(-Z) is the Gartner-Ellis limit for one consumer, and is given
by:
Using (26), (28), and (30), one can solve for z* that is found to be one
of the roots of a quadratic polynomial in z (hence, it is given in closed
form). Next, we compute c*. Let 'f7 ~ cdc = e/N = k/(nN). From (27),
z* can also be written in terms of the packet loss requirement (p, x) as
z* = _IO!p ~~. Substituting (28) and (30) in (26) and setting z = C we
arrive at the following equation:
(32)
Solving for c, we obtain a closed-form expression for the wireless EB c*
subject to a packet loss constraint:
c* = n2 - (A + f-t)2 (33)
2'f7~(f-t -,\ + n)
5. Numerical Results
In this section, we provide numerical results obtained based on the
previously presented analysis. A modulation scheme needs to be specified
16
in order to obtain the BER curve and the corresponding service ratios. In
our examples, we mostly focus on the binary phase-shift keying (BPSK)
and the differential phase-shift keying (DPSK) modulation schemes. For
BPSK modulation with coherent demodulation, the BER is given by:
(35)
In all of our examples, we let n = 424 and use Reed-Solomon (RS) code
for error correction. We also take E* = 10- 10 and 1m = 50 Hz. Unless
specified otherwise, we take p = 4 and T = 5 (hence k = 414 for the RS
code). The first two examples are intended to illustrate the partitioning
approach. The first example uses BPSK modulation. Following the al-
gorithm in Figure 4, we first obtain rN i=:::j 6.970 by solving (21) for rN.
From this value, we obtain rl = 0.396 and N = 2.9397, which we trun-
cate to three, i.e., the channel is represented by a 4-state FSMC. We then
recompute the values of rl and AI fJ,. The resulting partitioning, shown in
Table 1, is found adequate (i.e., satisfies the service-ratio criterion).
ReceivedSNR(r)
Figure 6. Service ratio (J(r) versus r for three modulation schemes (7 = 5).
model. Figure 7 depicts O(r) versus r for different values of T with BPSK
modulation. It is worth noting that the higher the capability of the FEC
code, the lower the expected number of states. This can be seen from the
sharper slopes of the service ratios in Figure 7. This can be explained
by the fact that the larger the value of T (stronger FEC code), the faster
O(r) approaches its asymptotic value, leading to a smaller value of N.
Figure 8 shows the solution for rl as a function of r N. This figure shows
that higher values of rN will result in smaller values for rl. The impact
of the choice of rN on N is shown in Figure 9 for different values of p (rl
fixed at 1.0). It can be seen that there exists an "optimal" value for rN
at which N is maximized, thus reflecting the channel characteristics more
accurately. Figure 10 shows the variation of N as a function of rl for
different values of p (rN fixed at 9.70). Thus, it is clear that the value of
the parameter N depends on the separation between the thresholds r N
and rl. Figure 11 shows the effect of p on N. It is observed that as p
increases, so does N, suggesting the possibility of using p as a means to
control N (since p can be controlled by adjusting the signal power at the
transmitter). Finally, from Figures 9 , 10, and 11 it is clear that a desired
value of the parameter N can be obtained by the appropriate choice of p
18
OB
i'
0.6
04
0.2
0
0 20
Figure 7. {}(r) versus r for BPSK with different error correction capabilities.
2.5,---,---,--.,..--.,---.,---,----,
1.5
'.
Figure 9. N versus rN for different values of p (rl = 1.0).
19
and the thresholds rl and r N that satisfy the service ratios as discussed
before. Next, we study the impact of our channel partitioning approach
However, in order to isolate the effect of N and gauge its impact on the
EB separately from other factors, we fix ,\ and f-l at their values in the
BPSK example (with N = 3). For the packet loss case, it is clear that for
a given QoS constraint the EB decreases dramatically with an increase
in N. This corroborates our intuition of the conservative nature of the
popular two-state Markov channel model. For example, by using four
states (N = 3) instead of two, c* subject to a PLR of 10- 6 is reduced by
almost 40% (and by 46% from the source peak rate). The reduction in the
EB can be explained by the fact that characterizing the channel with a
larger N reflects its error characteristics more accurately, and hence leads
to more efficient allocation. Figure 13 depicts the effective bandwidth as
a function of the PLR constraint p using different buffer sizes (x) with
N = 3. The figure shows that even with a small buffer, typical PLR
requirements (e.g., 10- 6 to 10- 3 ) can be guaranteed using an amount of
bandwidth that is less than the source peak rate. The significance of our
EB analysis is that it allows the network operator to decide beforehand
the amount of resources (buffer and bandwidth) needed to provide certain
QoS guarantees. A reduction in the per-connection allocated bandwidth
translates into an increase in the network capacity (measured in the num-
ber of concurrently active mobile users).
PLR
one may call the "optimal" FEC. Beyond r* the trend is reversed (Le.,
the overhead of FEC starts to outweigh its benefits).
I
~ 2000
!
~
~ 1500
~
1000
BufferSiz8
6. Conclusion
In this paper, we presented a new approach for partitioning the re-
ceived SNR range that enables tractable analysis of the packet loss and
delay performance over a time-varying wireless channel. This was done by
adapting the wireless channel to Mitra's producer-consumer fluid model,
which has known queueing performance. Our analysis exploited several
properties of a slowly-varying wireless channel, including its level-crossing
analysis and the Rayleigh distribution of the signal envelope. We provided
22
References
[2] H. Bischl and E. Lutz, "Packet error rate in the non-interleaved Rayleigh chan-
nel," IEEE Trans. Commun., vol. 43, pp. 1375-1382, 1995.
[3] H. Steffan, "Adaptive generative radio channel models," In Proceedings of the 5th
IEEE International Symposium on Personal, Indoor and Mobile Radio Commu-
nications., vol. 1, pp. 268-273, 1994.
[4] S. Lin and D. J. Costello, Error Coding: Fundamentals and Applications, Engle-
wood Cliffs, NJ: Printce Hall, 1984.
[5] W. C. Jakes, Microwave Mobile Communications, New York, NY: John Wiley &
Sons, INC., 1977.
[6] B. Vucetic, "An adaptive coding scheme for time-varying channels," IEEE Trans.
Commun., vol. 39, pp. 653-663, May 1991.
[7] M. Rice and S. B. Wicker, "Adaptive error control for slowly-varying channels,"
IEEE Trans. Commun., vol. 42, pp. 917-925,1994.
[8] H.S. Wang and N. Moayeri, "Finite-state Markov Channel- a useful model for
radio communication channels," IEEE Trans. Vech. technol., vol. 44, pp. 163-
171, 1995.
[9] Q. Zhang and Salem A. Kassam, "Finite-state Markov model for Rayleigh fading
channels," IEEE Trans. Commun., vol. 47, Nov. 1999.
[10] Y. L. Guan and L. F. Turner, "Generalized FSMC model for radio channels with
corrlated fading," IEEE Proc. Commun., vol. 146, April. 1999.
[11] J. Kim, and M. Krunz, "Bandwidth allocation in wireless networks with guaran-
teed packet-loss performance," IEEEIACM Transactions on Networking, vol. 8,
pp. 337-349, 2000.
[12] M. Hassan, M. Krunz, and I. Matta "Markov-based Channel Characterization
for Tractable Performance Analysis in Wireless Packet Networks," Submitted to
IEEE Transactions on Wireless Communications.
[13) D. Anick, D. Mitra, and M. M. Sondhi, "Stochastic theory of a data-handling
system with multiple sources," Bell Syst. Tech. J., vol. 61, pp. 1871-1894, Feb.
1982.
24
[14] D. Mitra, "Stochastic theory of a fluid model of producers and consumers coupled
by a buffer," Adv. Appl. Prob., vol. 20, pp. 646-676, 1988.
[16] A. I. Elwalid, and D.Mitra, "Statistical multiplexing with loss priorities in rate-
based congestion control of high-speed networks," IEEE Trans. Commun., vol. 42,
pp. 2989-3002,1994.
Key words: Voice over IP over Wireless, wireless Internet, statistical multiplexing.
1. INTRODUCTION
Key features, when designing voice services over IP for cellular radio
links, are spectrum efficiency and robustness to transmission errors. This
section provides a short overview ofVoIP over GERAN specific issues.
robust against errors to be used on the air interface. The CRTP (Compressed
RTP) algorithm [I], which is used for wireline VoIP, is not robust enough
for wireless links: if a compressed header is lost, the decompressor is not
able to reconstruct the subsequent headers (error propagation). Then, a single
packet error causes several consecutive lost packets (headers + voice
payloads). Some more adequate algorithms as ROHC (RObust Header
Compression) have been proposed [2]. ROHC is significantly less sensitive
to radio link errors thanks to repair mechanisms in the decompressor (no
error propagation) and reduces IPIUDPIRTP-packet header sizes down to
only 2 bytes most of the time [3]. In this paper, it is assumed that the
additional frame error rate due to header compression is negligible compared
to the other sources of errors further discussed in the following.
EGPRS is an evolution of GSM. Its air interface uses main basic physical
layer parameters of GSM (carrier spacing, TDMA frame and burst
structure), with the additional possibility of adaptive modulation: 8PSK
modulation is used instead of GMSK modulation when the radio link
conditions are favorable, which significantly increases the throughput.
Besides, different transmission rates are available. Depending on the chosen
modulation and coding scheme (MCS), the data rates range from 8.8 kbitls
(MCS-I) to 59.2 kbit/s (MCS-9) per time-slot (cf. Table I).
The channel coding schemes are derived from the same convolutional
code, having a rate of 113 and constraint length of 7 by applying different
puncturing schemes [4]. It should be noted that all information bits are
equally protected. The transmission on the radio interface in EGPRS is based
on radio blocks transported by 4 bursts (time-slots) in 4 consecutive TDMA
frames. The interleaving scheme is rectangular and limited to a depth of 4
frames. This structure in radio blocks enables a highly dynamic resource
28 A. Wautier, J Antoine, L. Husson, J Brouet , C. Thirouard
sharing on each PDCH: from one radio block to the next one, the resource
can be used by different users.
3. SYSTEM MODELS
TV 50, GSMJEFR
100
--TCHlEFR
--.\-MCS 3
--MCS6
10 1:-----4_-+-A-----It-----I ---- MCS 1+5
0,1
o 5 10 15 20 25 30 35
SNR(d8)
Fig. 2 - FER due to transmission channel
32 A. Wautier, J. Antoine, L. Husson, J. Brouet , C. Thirouard
Figure 2 illustrates the results obtained with a TU50 channel model with
different coding schemes encountered in GSM (TCHlEFR), in GERAN
(MCS3, MCS6), and with the coding scheme (MCS1+5) proposed in [5].
The Markov model represents a transmission including modulation, radio
channel, DFSE receiver structure with perfect channel impulse response
estimation. The Fritchman model has 2 bad states and 2 good states. Besides,
soft decisions at the output of the demodulator are considered [18, 19].
Moreover, perfect detection of erroneous frames is assumed.
While it is difficult to model the voice activity ofa single user, Weinstein
found that the number of active lines could be modeled by a continuous-time
birth-death process [20]. He showed that this model is quite valid when the
number of users is superior to 25. The parameters that govern the transition
rates are the mean talk-spurt duration a-I and the mean silence duration p-I.
The voice activity factor II is defined by the ratio of mean talk-spurt duration
to the sum of mean talk-spurt and mean silence duration:
11 = n. -1 -1
p +a
Typical values encountered in the literature for the voice activity factor
are 0.445 (obtained with a-I = 1.41 sand p-I = 1.74 s ) or 0.36 (obtained
with a-I = 0.96 sand p-I = 1.69 s).
If btlt) is the probability of having i active lines at time t assuming that
we have} admitted voice communications, the steady-state probability of
having i active lines among} admitted lines is given by:
bijj -- (i}
i i (1- 11 )j-i fior 0 <
- l. <
- )
.
The size of the pool and the number of pools per TRx will impact the
statistical multiplexing capacity gain. Table 3 summarizes the number of
pools p of size c per TRx (p x c) for different codecs and physical link
configurations versus the mobile station (MS) multi-slot capability.
MS multi-slot capability I 2 4 8
GSMFRorEFR MCS3 8xl 4x2 2x4 lx8
MCS6 8x2 4x4 2x8 lx16
GSMHR MCS3 8x2 4x4 2x8 lx16
MCS6 8x4 4x8 2x16 lx32
Table 3 - Possible values for the number serving channels per TRx (pxc)
34 A. Wautier, J. Antoine, L. Husson, J. Brouet , C. Thirouard
The packet loss rate can be analytically obtained for a given number of
active users denoted by i assuming j admitted users. The considered voice
activity model is the one described by Weinstein [20]. A packet is lost when
the buffer is full (the actual number of packets in the buffer is q = m), this
can of course occur only when i is greater than c. Then, the packet loss rate
can be expressed as [5]:
j .
Considering the traffic model seen in 3.3 and assuming that the
maximum number of admitted users is N, the relationship between the
blocking probability and the offered traffic p is given by the classical Erlang
B formula [21]:
pN
P(N)= N!
I-k!
N pk
k:O
The worst case for the packet loss rate is obtained for j=N, which is given
by p ~N). Voice activity model can be combined with the traffic model: the
voice activity Markov chain can be viewed as a sub-chain of the traffic chain
[6]. Then, assuming that the two models are independent, the mean packet
loss rate is given by:
pj
I
N
Pd = Pd(J)P(J) where P(J·) = N
j! k
I~
j:c
k:O k!
The criterion to evaluate the capacity (possible offered traffic) must take
into account not only the blocking probability threshold of 2% but also a
criterion about the packet loss probability, which can be either a worst
packet loss rate threshold (1 % for example) or a packet loss rate threshold
for a given percentage of time (e. g. a threshold of 1% for 95% of time). In
the latter case, the distribution of the packet loss has to be considered.
Capacity analysis of VolP over GERAN with statistical multiplexing 35
Two models are used to evaluate the quality of speech: the PESQ model
and the E-model.
The PESQ method (Perceptual Evaluation of Speech Quality) is dedicated
to end-to-end speech quality assessment of narrow band telephone networks
and speech codecs [12]. The original speech samples and the same but
degraded samples passed through a communication system are compared
using a psycho-acoustic model of the human ear. This method is also valid
for communication systems introducing distortions (e.g. time misalignment,
transmission errors, ... ). It is thus relevant for assessing the impact of the
radio link on voice quality perception.
The PESQ delivers MOS scores which is convenient to make the link with
other methods. In particular, the PESQ permits to calibrate some factors used
in the E-model and both models are thus complementary.
The E-model is a parametric model, which assumes that the degradations
due to different factors are cumulative and therefore studied separately [13].
The quality measurement R uses a scale between 0 (poor) and 100 (good)
which is linked to the MOS scale by the following transformation rule (it
should be noted that the degradations are only cumulative in the R scale)
[13]:
In this paper, we consider two terms in the quality evaluation: ldd and Ie.
/)e-~
Degradation due to equipments (Codec. FER) Degradation due to delay
The parameter ldd depends on the transmission delay Ta due to the size of
the buffer and to the interleaving depth of the coding scheme. Figure 3-a
plots ldd whose expression is [13]:
The parameter Ie depends on the intrinsic quality of the codec and on the
sensitivity to the FER (which includes the processing made to restore lost
frames). This degradation does not come from a closed-form formula but is
determined through listening tests or objective methods like the PESQ
method. For example, the GSM EFR codec has an intrinsic quality of 4,32 in
the MOS scale. In this paper, we used the PESQ method to determine Ie.
Figure 3-b gives Ie in case of uniformly distributed packet loss for GSM EFR
codec obtained with PESQ simulation and it is compared to the normalized
MOS evaluation [13]. This actually reflects the packet dropping due to
statistical multiplexing.
30 30
ri--Mosnonn V
25 25 _ _ PESQnote
20
/
V/V
20
J
15
~
15
~V
~
10 10
~~
~
o 100 200 300 400 500 0,1 10
Delay (ms) FER(%)
(a) (b)
Fig. 3 - Influence of delay and FER on speech quality degradation with E-model
E(talk]= 1.415 ; E[silence1= 1.745 ; c=16; Delay between 0 and 6 speech frames
5%
4%
~
j 3% I ................. , ..........................,..... +. . . . . . + ;············I+IIIH,r.····j
~
~
as 2%
t
1% I·························,··························· ........•........ ,············//AijH > ••..••••••.•••.•..• ~
oL---~---=~~~~~-L--~
16 18 20 22 24 26 28 30 32 34
number of users J
~
~ 80% ~
packet loss rate =1% ~ 200%~
·5
"
" 60% ~
100%'
40%
O%:L'_ _ _ _ _ _ _==~~
20%0 5 10 15 20 25 30 35 40 45 50 0.2 0.3 0.4 0.5 0.6 0.7 0.8
C
Speech activity factor 11
Fig. 5 - Influence of the size of the channel Fig. 6 - Influence of the speech activity
pool on capacity gain for a given FER factor on the capacity gain
.............................................................................................
The delay and the packet loss rate are two parameters that induce
different effects on the perceived quality. The resulting degradations are
separately evaluated in the E-model (refer to figure 3). However, there is a
dependency between delay and FER. This relationship is illustrated in
figure 7, which confirms that the packet loss rate decreases as the delay is
increased. Considering both influences on the quality measurement leads to a
tradeoff as shown in figure 8 for the EFR codec.
70 L -_ _---L_ _ _--L.._-'-:-:-_.L..._,.-:----'
o m ~ ~ M m m ~ ~
The radio link behavior impact on the perceived speech quality has been
evaluated by means of simulation. The simulator models a complete
transmission chain comprising voice frame generation, coding, channel error
model, decoding, and voice frame reconstruction. The channel model that is
included in the simulator is a two-layer model replacing the modulation-
channel-demodulation as described in 3.1. The output parameters of the
simulator are the FER at the output of the decoder and the voice quality
measure obtained with the PESQ method. The MOS obtained with PESQ
method is converted in the R scale to be further exploited in the E-model
through Ie factor. Speech quality evaluation with E-model is illustrated in
figure 9 for the GSM EFR codec for three different MCS as a function of the
SNR.
TUSO, GSMlEFR + lost frames processing
100
90
80
..IV 0
.1 -;
70
I . . . I-r"'"
...~
0
60
I / /'
/'
,
50
~=: ./ I /
,
40 --TCHlEFRR
/ I I / --MCS3 R
/ /
30
--MCS6R
"/
20
10 I' -.l-MCS9 R
--MCS1+5R
0
o 10 15 20 25 30 35
SNR(dB)
Fig. 9 - Influence of radio link performance on speech quality (no statistical multiplexing)
5. CONCLUSION
Capacity evaluation for voice service over IP over GERAN packet radio
bearers based on speech quality estimated through E-model and PESQ
methods, is presented. After an overview of main aspects of VoIP over
GERAN, we have evaluated the impact of statistical multiplexing on
capacity by combining closed-form analytical studies and simulations.
Substantial capacity gain can be obtained through dimensioning of
system parameters (e.g. buffer size, frequency reuse factor, ... ). Besides,
straightforward transmission of voice frames on PDCH can lead to high
requirements in terms of SNR. It has been shown however, that those
requirements can be easily mitigated, resulting in more typical SNR target
[5].
An exhaustive study of VoIP over GERAN should ideally consider
second order impacts resulting from header compression mechanisms as well
as the associated signaling channel overheads. Finally, system complexity
should be addressed and compared with more conventional solutions.
References
[1] "RFC 2508, Compressing IPIUDP/RTP Headers for Low-Speed Serial
Links", IETF, February 1999.
[2] "RFC 3095, RObust Header Compression (ROHC): Framework and four
profiles: RTP, UDP, ESP, and uncompressed", IETF, July 2001.
[3] L. Larzon et al. "Efficient transport of voice over IP over cellular links",
Proceedings ofPIMRC'OO, London, Sept. 2000.
[4] "3GPP TS 05.30: Channel coding", v. 8.6.1, Release 1999, January 2001.
42 A. Wautier, J Antoine, L. Husson, J Brouet , C. Thirouard
1. INTRODUCTION
L2
The radio interface of the UTRA is layered into three protocol layers: the
Physical Layer (Ll), the Data link Layer (L2) and the Network Layer (L3).
Additionally, the layer 2 is split into two sub-layers, the Radio Link Control (RLC)
and the Medium Access Control (MAC). On the other hand, the RLC and layer 3
protocols are partitioned in two planes, namely the User plane and the Control plane.
In the Control plane, Layer 3 is partitioned into sub layers where only the lowest
sublayer, denoted as Radio Resource Control (RRC), terminates in the UTRAN, as
Figure 1 shows.
Connections between RRC and MAC as well as RRC and Ll provide local
inter-layer control services and allow the RRC to control the configuration of the
lower layers. In the MAC layer, logical channels are mapped to transport channels.
A transport channel defmes the way in which traffic from logical channels is
processed and sent to the physical layer. The smallest entity of traffic that can be
transmitted through a transport channel is a Transport Block (TB). Once in a certain
period of time, called Transmission Time Interval (TTl), a given number of TB will
be delivered to the physical layer in order to introduce some coding characteristics,
interleaving and rate matching to the radio frame. The set of specific attributes are
referred as the Transport Format (TF) of the considered transport channel. Note that
the different number of TB transmitted in a TTl indicates that different bit rates are
associated to different TF. As the UE may have more than one transport channel
simultaneously, the Transport Format Combination (TFC) refers to the selected
combination of TF. The network assigns a list of allowed TFC to be used by the UE
in what is referred as Transport Format Combination Set (TFCS).
45
It is worth mentioning that for the optimisation of the radio interface utilisation,
RRM functions should consider the differences among the different services, not
only in terms of QoS requirements but also in terms of the nature of the offered
traffic, bit rates, etc. The RRM functions include:
1. Admission control: it controls requests for setup and reconfiguration of radio
bearers.
2. Congestion control: it faces situations in which the system has reached a
congestion status and therefore the QoS guarantees are at risk due to the
evolution of system dynamics (mobility aspects, increase in interference, etc.).
3. Mechanisms for the management of transmission parameters: are devoted to
decide the suitable radio transmission parameters for each connection (i.e. TF,
target quality, power, etc.).
4. Code management: for the downlink it is devoted to manage the OVSF code
tree used to allocate physical channel orthogonality among different users.
Within the UMTS architecture, RRM algorithms will be carried out in the Radio
Network Controller (RNC). Decisions taken by RRM algorithms are executed
through Radio Bearer Control Procedures (a subset of Radio Resource Control
Procedures) such as [3]:
1. Radio Bearer Set-up.
2. Physical Channel Reconfiguration.
3. Transport Channel Reconfiguration.
3GPP has provided a high degree of flexibility to carry out the RRM functions,
so that the parameters that can be managed are mainly:
1. TFCS (Transport Format Combination Set), which is network controlled
and used for Admission Control and Congestion Control.
2. TFC (Transport Format Combination), which in the case of the uplink is
controlled by the UE-MAC
3. Power, as the fundamental physical parameter that must be set according to
a certain quality target (defined in terms of a SIRtarget) and taking into
consideration the spreading factor used and the impact of all other users in
the system and their respective quality targets.
4. OVSF (Orthogonal Variable Spreading Factor) code
In the above framework, this paper focuses on the admission control and the
mechanisms for the management of transmission parameters for conversational and
interactive services carried out at UE-MAC in the uplink direction. It is worth
mentioning that the problem of QoS provisioning for multimedia traffic has gained
interest in the literature in recent years, as the problem arises in the context of 2.5G
and 3G systems and is not present in 2G systems. Thus, Naghshineh and Acampora
[4] introduced resource sharing schemes for QoS guarantee to different service
classes in microcellular networks. Akyildiz et al. [5] proposed the so-called
WISPER protocol, scheduling the transmissions according to their BER
requirements. Das et al. [6] developed a general framework for QoS provisioning by
combining call admission control, channel reservation, bandwidth reservation and
bandwidth compaction. Lately, Dixit et al. [7] among others have discussed the
evolution scenarios from 2G to 3G networks and the QoS network architecture
46
proposal by 3GPP for UMTS. For the decentralized uplink RRM component,
proposals such as the ones presented in [8] could be adapted to the VTRA-FDD
framework. In this respect, few studies aligned to the 3GPP specifications are
available in the open literature [9-11]. For the admission control several schemes
have been suggested for the uplink [12, 13] under different conditions and at a lower
extent for the downlink [14]. More recently, Ho et a1. [15] have built mathematical
models for various call admission schemes and have proposed an effective linear
programming technique for searching a better admission control scheme. The
admission approach presented in this paper is innovative in the sense that the
admission procedure in the uplink is related to the decentralised algorithm applied at
VE-MAC level and the relevance to take this fact into account is pointed out.
The rest of the paper is organised as follows. Section 2 details the uplink RRM
approach by proposing two different VE-MAC algorithms and a statistical-based
admission control strategy. Section 3 details the simulation model used to evaluate
the strategies through system level simulation in Section 4, where some basic
assumptions concerning congestion control are introduced. Finally, Section 5
summarises the obtained results.
2. UPLINK RRM
A B A B
<:==:c> <!===:>
:'
.--~---t-··"-·/
OFFERED LOAD
Within a CDMA cell, all users share the common bandwidth and each new
connection increases the interference level of other connections, affecting their
quality expressed in terms of a certain (E,/N,). For n users transmitting
simultaneously at a given cell, the following inequality must be satisfied [18]:
(2)
where Pi is the k-th user received power at the base station, SFi is the i-th user
spreading factor, PN is the thermal noise power, X is the intercell interference and
(E,/N,)i stands for the i-th user requirement. P R is the total received own-cell power
at the base station. Implicitly in the above inequalities a certain received power level
is assumed in each case:
i=1..n (3)
(4)
49
[~J
PR ~ _ _ _ _ _N_O---'-i_ _ => PR =---------'----
[~J (5)
1 1
1-2::--- 1-2::---
n n
SF;
i=1 SF;
i=1
+1 +1
Claiming in (5) for the inherent positivity of PR (i.e. PR>O) leads to:
n 1 (6)
I---<1
i=1 SF; + 1
Claiming in (5) for the inherent positivity of % (i.e. %>0) leads to:
[!:J,
Claiming in (5) for the inherent positivity of PN (i.e. PN >0) leads to:
:J~
(8)
( 1+ SF;l + 1< 1
(~:l
The later expression is commonly known as the load factor [19]. The load factor
measures the theoretical spectral efficiency ofa W-CDMA cell:
(9)
50
Notice that 7]<1 is equivalent to claim that PN>O and so 7]<1 is the same
expression as (8). Introducing the definition of the load factor 17, (3) can be
expressed as:
P _1_
(PN + X +PR ) N 1-1"] i=l..n (10)
P> =---~
,- SF SF
---=-'-+1 j +1
where it can be observed that as the load factor increases the power demands also
increase. Consequently, and due to the limited power available at mobile terminals
and also for efficiency reasons the cell load factor must be controlled. Admission
control is one of the RRM strategies devoted to achieve such an objective.
The admission control procedure is used to decide whether to accept or reject a
new connection depending on the interference (or load) it adds to the existing
connections. Therefore, it is responsible for deciding whether a new RAB (Radio
Access Bearer) can be set-up and which is its allowed TFCS. Admission control
principles make use of the load factor and the estimate of the load increase that the
establishment of the bearer request would cause in the radio network. From the
implementation point of view, admission control policies can be divided into
modeling-based and measurement-based policies [20]. In case the air interface load
estimation is based on measurements and assuming that K users are already admitted
in the system, the (K+ 1)th request should verify:
< . _ PR +X l+l
11 + ~11 - 11 max With 11 - ---'-'---'-'-- and PR (11)
PR + X + PN ~ll = --S-'F---"----
_ _---"K:...c+-=--l_ _ +1
VK+l.(~b) o K+l
VK+l being the activity factor of the (K + l)th traffic source. In the case of the voice
service this factor is typically set to 0.67. For interactive services, like www surfing,
this factor should be estimated on a service by service basis.
Capacity and coverage are closely related in W -CDMA networks, and therefore
both must be considered simultaneously. In the measurement based approach llrnax is
obtained from radio network planning so that coverage can be maintained. The
coverage problem is directly related to the power availability, so that the power
demands deriving from the system load level should be in accordance with the
planned coverage. So, it must be satisfied that the required transmitted power will be
lower than P Tmax allowed and high enough to be able to get the required (EbINo)
target even at the cell edge:
51
P _1_
P =L (d) (pN + X + PR ) =L (d) N 1-11 i=1..n (12)
T., p' SF. p , SF
----"-+ 1 ; +1
( ~)
No ;
P 1
P . =L (R) N 1 -11 max i=1..n (13)
T.max p SF.
----'-,- + 1
( ~)
No ;
PT.; being the power transmitted by the i-th user, Lp (d;) the path loss (including
shadowing effects) at distance d; R the cell radii and T]max the maximum allowable
load factor for assuring coverage. The term 1I( 1 -17 ) is known as the interference
margin.
In case the air interface load is estimated in statistical terms it is the cell
throughput which is maintained and cell breathing effects may arise due to the fact
that intercell interference can not be directly and precisely included. For the
statistical-based approach and assuming that K users are already admitted in the
system, the (K +1)th request should verify:
K 1 1
+ (1 + f) (14)
(1+ f)I :<;;11 max
;=1 SF; SFK+1
+1 +1
v, {!:} VK+1(!b )
o K+1
where other-cell interference power is modeled as a fraction of the own-cell received
power (X=jxPR). According to (14) different admission strategies arise by balancing
the following parameters:
>- The spreading factor: by setting SFj as an estimated average value the user
will adopt along its connection time the assumed load will be closer to the
real situation at the expense of relying on the statistical traffic multiplexing.
In turns, considering SFj as the lowest SF in the defmed RAB covers the
worst case at the expense of overestimating the impact of every individual
user and, consequently, reducing the capacity.
>- The activity factor of the traffic source: by setting V; <1 the admission
procedure can be closer to the real situation of discontinuous activity
(typical in interactive-like services) at the expense of relying on the
statistical traffic multiplexing. In turns, V; =1 covers the worst case at the
52
TBmax being the maximum number of Transport Blocks that can be transmitted per
TTl and TBsize being the number of bits per Transport Block.
where SCr(n) is the Service Credit for TTI=n, SCr(n-l) is the Service Credit in the
previous TTl, GuaranteedJate is the number of bits per TTl that would be
transmitted at the guaranteed bit rate, TB_size is the number of bits of the Transport
Block for the considered RAB, Transmitted_TB(n-l) is the number of successfully
transmitted Transport Blocks in the previous TTL
The quotient GuaranteedJatelTB _size reflects the mean number of transport
blocks that should be transmitted per TTl in order to keep the guaranteed mean bit
rate. As a result, SCr(n) is a measure of the number of Transport Blocks that the
53
connection should transmit in the current TTl to keep the guaranteed bit rate. For
example, if TB_size=240 bits, GuaranteedJate=24 Kb/s, and TTI=20 rns, the UE
adds 2 service credits each TTL
Then, assume that in the buffer there are Lb bits, the number Transport Blocks to
be transmitted in the current TTI=n would be:
The radio access bearer considered for supporting the interactive service has a
maximum bit rate of 64 Kbps in the uplink and an associated 3.4 Kbps signalling
radio bearer [21]. The radio access bearer selected for videophone service has a
constant bit rate of 64 Kbps when transmitting [21]. TB error rate target is 0.5%.
Possible transport formats are detailed in Table 1.
The interactive traffic model considers the generation of activity periods (i.e.
pages for www browsing), where several information packets are generated, and a
certain thinking time between them, reflecting the service interactivity. The specific
parameters are: average thinking time between pages 30 s, average number of packet
arrivals per page: 25, number of bytes per packet: average 366 bytes, maximum
6000 bytes (truncated Pareto distribution), time between packet arrivals: average
0.125 s, exponential distribution. The videophone traffic model is a constant bit rate
source of 64 Kbps with average duration 120s. As the interest of the present paper in
what admission control concerns is on the statistical terms in (11), the simulation
model includes a cell with radii 0.5 km and intercell interference is represented by
f=0.6. Physical layer performance, including the rate 1/3 turbo code effect and the
1500 Hz closed loop power control, is taken from [22] to feed the system level
simulator presented here with BLER (BLock Error Rate) statistics. The mobility
54
model and propagation models are defined in [23], taking a mobile speed of 50 km/h
and a standard deviation for shadowing fading of 10 dB.
4.- RESULTS
MR
SCr24 0.9 r - - - - - -- - - -- - - -
G . r - - - -- - - - - ---- 0.8 t - - - - - - -- - --
el.lS 0.7t - - - -- - - -- - -
OJ o.&t----- - - -- - -
02' 0.5t - - - - - -- - - --
0.• t - - -- - - - -- - -
0.150 0.3 t_- - - - -- - - - -
G.' O~ t_----------
0.050 0.1 t - - - - -- - -
O~. .~~--. .~-
IF. TF. T.2 TFJ TF ..
According to (13), Figure 6 plots the maximum cell radii for a 95% coverage
probability as a function of the load factor for the considered services with a cr= 10
55
dB shadow fading and unity antenna gains. Thus, for a cell range of 500m, the load
factor must be below 75%.
Admission probabilities for videophone service with 11 max =0.75 are shown in
Table 3. Since videophone is of constant bit rate nature, the admission procedure is
quite easy to handle. For 15 users in the cell, the system is still below the planned
load factor, all users can be admitted and the performance is good, as shown in
Figure 7 and Figure 8. In particular, Figure 7 shows the power limitation probability
(i.e. the probability that the required transmitted power is above the maximum
value) and Figure 8 the TB error rate also as a function of the distance to the cell
site. If the offered load is above the planned load factor, as it is the case for 30 users,
the admission procedure is able to assure the system stability, planned coverage and
planned quality by rejecting connection requests. Figures 7 and 8 show a little
increase in the power limitation probability as well as the TB error rate due to the
increase in the load factor. In turns, Figure 9 shows the load factor distribution in the
system.
c: 0,05
0
E :g~ 0,04
. .
0,03 __ 15 users
~ ftI
,g
0,02 __ 30 users
G) 0
~ CL
0,01
0
CL
0
o II) 0 II) 0 II) 0
I'- II) N 0 I'- II)
~ N C") ('I') 'V
Distance to cell site (m)
-
a~4
GI 3 l
1--15
....
1U
0:::
2
/, users
w 1
--30 users
0
/~
...-..-
10
l- 0
\)
'\~ ,,'0\) ~ s::,\) (I..~ 0\)
'1,. ~ '!l l>i
Distance to cell site (m)
40~--------------------------,
~ 30+--------------T~----------~
~ 20+-------------~~----------~
.c
£ 10 +------------"7....------'1,.----------1
o N M ro
a a a a a a a a a
~ ~ ~ ~ ~ ~
Load factor
Admi'
T,ahie3. SSlOn probabT'
I It1es fIor VI'deoplhone serviceo
Number of videophone users Admission
probability
15 1
20 0.98
25 0.79
30 0.59
For the interactive service the situation is not so easy to handle because of two
dynamic issues affecting the system behaviour that are difficult to predict: the
statistical traffic multiplexing (the interactive service is of discontinuous nature and,
consequently, the number of simultaneous users in a given frame in principle is not
known in advance) and the TF used in uplink transmissions (it is decided in a
decentralized way by UE-MAC and, consequently, the set of SF used by
simultaneous users in a given frame in principle is not known in advance). Table 4
shows the admission probabilities for different values of the admission TF
57
(equivalent to SF) and 11 max for both MR and SCr strategies. The activity factor is
assumed to be the average value coming from the traffic model. The criterion for
considering the system under a congested situation is when (18) holds for more than
90 out of 100 consecutive frames, revealing that the CDMA capacity has been
overcome. Note that depending on the specific congestion detection and congestion
resolution algorithms, the system could continue operating under normal conditions
or not and the interest of the present criterion is only for establishing a basis for
comparison purposes since we are not dealing with congestion control algorithms in
this paper.
n
1 (18)
(1+ /)"L >l1th
SF;
(~)
;;1
+1
No ;
It can be observed from Table 4 that for a proper admission procedure the
characteristics of the decentralized algorithm being applied at DE-MAC layer should
be taken into account. For example, if TF2 and 11 max =0.75 are considered in the
admission phase for MR strategy, and since the dynamic behavior of this algorithm
tends to use TF4 in most cases, the system enters in congestion with 500 users
because the admission is too soft. In turns, for SCr the TF considered for admission
purposes is much better adjusted to the real dynamic value, so that admission allows
for more than 550 users to enter in the system while maintaining a controlled
performance. On the other hand, if TF4 is considered for admission purposes,
congestion is avoided because from the transmission rate point of view the worst
case is considered and from the traffic multiplexing point of view 11 max =0.75 is low
enough to absorb traffic fluctuations without causing congestion. Nevertheless, for
SCr strategy this is not so suitable because the admission is too strict. It is worth
noting that the value for 11 max eventually allows for a softer or stricter admission as
shown in the example in Table 4, where increasing the value up to 11 max =0.9
58
improves the percentage of admitted users for SCr strategy compared to the TF4 and
11 max =0.75 case. For this later case, Figure 10 plots the statistical load factor
distribution, showing that it tends to be quite low because of the still strict admission
procedure. For 550 wwwusers, 2% of the requests are already rejected (see Table 4)
while the performance in the system is still very good. As a matter of fact, Figure
11 shows that the power limitation probability as a function of the distance to the
cell site (i.e., the probability that a given user requires more power than the
maximum allowed for achieving the target EbINo) is within the coverage probability
design even at the cell edge. Also, Figure 12 plots the average packet delay again as
a function of the distance to the cell site, and reveals that no performance
degradations are observed as one moves far from the cell site.
100
10
~
~ 0.'
:;;
"
~
0,01
Q. 0,001
0.0001
0,00001
,
'" "'. ",'!- ",,,
"'.'" ",'? "'.
~
Load factor
l>
~ 0,025 ~---';";"'---:--...."......,..,.___;;"'::;:"";''''';'''--4-I
!
!!
0,01 ~"";:,~=-=,..,-~.,,...;,.....,..,=,::.,-..-...,--'-..r---"-;""'--l
t 0.005~;""'~~~'--~;""'-~--:-:~~--:-:~___--l
o
Q.
~ ~ &# $ # $ # ~
Olsta nee to ce II site (m)
~ # $ # # # $ # ~
DISIa nee to cell site (m)
5. CONCLUSIONS
3G will offer different QoS guarantees and an optimization of capacity in the air
interface by means of efficient RRM algorithms, which should consider the
differences among the different services, not only in terms of QoS requirements but
also in terms of the nature of the offered traffic, bit rates, etc. In the framework of
VTRA-FDD, this paper has focused on the admission control and the mechanisms
for the management of transmission parameters for conversational and interactive
services in uplink direction. Results for VE-MAC strategies show that for SCr
strategy most of the time TFI and TF2 are used while for MR strategy most of the
time TF4 is used This different behaviour of the VE-MAC algorithms impacts on
the admission control process, which should take this fact into account for avoiding
either too strict or too soft policies.
6. ACKNOWLEDGMENTS
This work is part of the ARROWS project, partially funded by the European
Commission under the 1ST framework (1ST 2000-25133) and by the Spanish
Research Council under grant TIC2000-2813-CE.
7. REFERENCES
[I] 3GPP TS 25.211, "Physical channels and mapping of transport channels onto physical
channels (FOO)"
[2] 3GPP TR 25.922 v4.0.0, "Radio resource management strategies"
[3] 3GPP TS 25.331 v4.0.0, "RRC protocol specification"
60
[4] M. Naghshineh, A. S. Acampora, "Design and control of micro cellular networks with QoS
provisioning for data traffic", Wireless Networks 3 (1997) , pp. 249-256.
[5] I. F. Akyldiz, D. A. Levine, I. Joe, "A Slotted CDMA Protocol with BER Scheduling for
Wireless Multimedia Networks", IEEE/ACM Transactions On Networking, Vol. 7, No2,
April 1999, pp. 146-158.
[6] S. K. Das et ai, "A call admission and control scheme for QoS provisioning in next
generation wireless networks", Wireless Networks 6 (2000) ,pp. 17-30.
[7] S. Dixit, Y. Guo, Z. Antoniou, "Resorce Management and Quality of Service in Third
Generation Wireless Networks", IEEE Communications Magazine, Vol. 39, No.2, February
2001.
[8] O. Sallent, R. Agusti, "A Proposal for an Adaptive S-ALOHA Access System for a
Mobile CDMA Environment", IEEE Transactions on Vehicular Technology, Vol. 47, n° 3,
August 1998.
[9] K. Dimou, P. Godlewski, "MAC Scheduling for Uplink Transmission in UMTS W-
CDMA", Proceeding of the IEEE Vehicular Technology Conference, VTC 2001 Spring.
[10] R. De Bernardi et aI., "Load Control Strategies for Mixed Services in WCDMA",
Proceedings of the Vehicular Technology Conference 2000, VTC'OO, pp. 825-829.
[II] W. Rave et aI., "Evaluation of Load Control Strategies in an UTRAlFDD Network",
Proceedings of the Vehicular Technology Conference 200 I, VTC'O I, pp. 2710-2714.
[12] H. Holma, J. Laakso, "Uplink Admision Control and Soft Capacity with MUD in
CDMA", Proceedings of Vehicular Technology Conference, VTC'99 Fall.
[13] Z. Lui, M. El Zarki, "SIR Based Call Admission Control for DS-CDMA Cellular
Systems", IEEE Journal on Selected Areas in Communications, Vol. 12, 1994.
[14] J. Knutsson et aI., "Downlink Admission Control Strategies for CDMA Systems in a
Manhattan Environment", Proceedings of Vehicular Technology Conference, VTC'98.
[15] C. J. Ho et aI., "On Call Admision Control in DS/CDMA Cellular Networks", IEEE
Transactions on Vehicular Technology, Vol. 50, n° 6, November 2001.
[16] R. K. Morrow, J.S. Lehnert "Packet Throughput in Slotted ALOHA DS/SSMA Radio
Systems with Random Signature Sequences", IEEE Transactions on Communications, Vol.
Com. 40, No.4, July 1992, pp. 1223-1230.
[17] N. Abramson, "Multiple Access in Wireless Digital Networks", Proceedings of the IEEE,
Vol. 82, No.9, September 1994, pp. 1360-1369.
[18] O. Sallent, J. Perez-Romero, F. Casadevall, R. Agusti, "An Emulator Framework for a
New Radio Resource Management for QoS Guaranteed Services in W-CDMA Systems",
IEEE Journal on Selected Areas in Communications, Vol. 19, No. 10, October 2001.
[19] H. Holma, A. Toskala (editors), W-CDMAfor UMTS, John Wiley and Sons, 2000.
[20] V. Phan-Van, S. Glisic, "Radio Resource Management in CDMA Cellular Segments of
Multimedia Wireless IP Networks", WPMC'OI Conference Proceedings.
[21] 3G TS 34.108 v.3.2.0, "Common Test Environment for User Equipment. Conformance
Testing"
[22] J. Olmos, S. Ruiz, "UTRA-FDD Link Level Simulator for the ARROWS Project",
IST'OI Conference Proceedings, pp. 782-787.
[23] 3GPP TR 25.942 v.2.1.3, "RF System Scenarios"
Rate and Power Adaptation for Downlink Shared
Channel in WCDMA
1. INTRODUCTION
code, several lower bit rate users can be allocated lower rate physical
downlink shared channels on a code multiplexing principle. Downlink shared
channels allocated to users on different radio frames may have different
spreading factors. Each PDSCH is associated with a downlink Dedicated
Physical CHannel (DPCH) to achieve power control and signaling. All
relevant Layer 1 control information is transmitted on the associated
Dedicated Physical Control Channel (DPCCH); i.e. the PDSCH does not
carry any Layer 1 information.
In the downlink, each base station has a maximum power budget that is to
be shared among users belonging to the different QoS classes. A portion of
this base station power is allocated to common control channels such as base
station specific beacon channel and pilot channel. The remaining power is
available for traffic or information channels.
The users in a cell are allocated power levels according to their QoS
requirements and their relative locations. The number of simultaneous users
served by a base station is limited by its maximum radiated power (Pmax).
The received average signal-to-interference (SIR) by a given mobile can be
expressed as:
SIR = W. ~ (1)
R Idownlink
Where ~ is the received average signal power, Idownlink is the total
downlink interference, W is the chip rate and R is the transmission rate.
Since the downlink codes are not completely orthogonal within a cell due to
multipath propagation, the total downlink interference is given as:
I intra is the interference from within the home cell, I inter is the
(3)
64
(4)
If there are m services and N; users of service class i in the cell, the
following condition must be satisfied to ensure proper operation:
(5)
I
FIFO Priority 2
Scheduling Queue
I I I I
r+8
~ II ~ DSCH
To admit interactive users, waiting in a CAC queue, the system checks the
associated control channel (DPCCH) power requirements and the mean
transfer delays of the on going connections. Waiting users are not transferred
to the scheduling queue if the delays grow excessively. The average delays
are used as an indirect way to sense high traffic loads and prevent new
requests from entering the serving queue. Upon admission, the users are
transferred to the scheduling queue and allocated a dedicated control channel
for signaling and control purposes.
As indicated, the scheduling queue monitors message delays to assist
CAe. Active connections, in the scheduling queue, are served according to
the Earliest Deadline First policy. The interactive class users are scheduled
only if their link quality is acceptable and higher priority services link quality
is respected as well. By giving precedence to high priority real time users,
only the remaining radio resources are allocated to non real time packet users.
In downlink WCDMA, interference greatly depends on user position.
When a base station operates at low load or the mobile station channel is in
good condition, smaller processing gain and higher transmit power can be
applied to the interactive users. On the contrary, if the base station is
operating at high load or the mobile station experiences poor radio link
quality, the base station decreases the rates of interactive users to stabilize the
system.
66
BEGIN
P minji : Required power at minimum rate for mobile j and service class i
P max,i : Maximum allowed power for mobiles of service class i
R min : Minimum rate for NRT service class i = 32 Kbps
R max,i : Maximum rate for NRT service class i = 256 Kbps
Figure 2 shows the flowchart for power and rate adaptation over a
scheduling cycle (a frame). Users are first sorted according to channel
condition between mobile and base station. The channel state information is
provided by the associated DCCHs. The base station can also estimate the
DSCH power level through the associated DCCHs power levels when
deciding transmission rates for mobile stations on DSCHs. Users are then
selected according to the Earliest Deadline First priority policy. The deadline
for each user is calculated according to minimum acceptable throughput and
packet size, assuming all NRT users have the same maximum transmission
rate capability.
NRT users are only scheduled for downlink packets transmissions if their
required transmission power at minimum allowed rate is less than the power
budget for each user. In addition, the base station is not operating at maximum
threshold power. Starting from the maximum allowed rate, transmission rate
67
is decreased to the next lower level if the required mobile transmission power
is greater than the maximum mobile power budget limits for both mobile and
base. In other words, each request is first checked for feasible allocation at
minimum rate. Once this first test is passed, the other rates are checked
sequentially starting for the highest possible rate. The algorithm can be
improved of course but the paper objective is to simply assess the benefits of
rate adaptation combined with CAC and scheduling.
Once the scheduler has decided about user rate and power, the availability
of the PDSCH is checked. If a shared channel of corresponding rate is
available, the mobile gets the reservation for transmission of a Service PDU.
Otherwise, the mobile's packets simply wait in the scheduling queue for the
next scheduling instant. In this way mobiles are segregated on shared
channels of different rates depending upon their locations and channel
conditions.
Rate matching is achieved by mapping NRT users on PDSCHs giving bit
rate of 32, 64, 128 and 256 Kbps at RLC payload. Packets are segmented into
fixed size transport block (RLC PDU) of 320 bits (uncoded). The scheduler
determines the data rate and the number of transport block (RLC PDU) to be
transmitted according to the transmission rate. Therefore, rate adaptation
results in transmission of 1,2,4 or 8 RLC PDUs within a frame.
3. SIMULATION MODEL
4. SIMULATION RESULTS
Simulation results are reported in Fig. 3 through Fig. 11. Fixed rate
scheduling is used for comparison and serves as a reference. Higher priority is
assigned to the RT type service at 64 Kbps behaving like a conversational
class. The RT service load is held constant at 4 Erlangs. RT traffic flows are
transmitted on dedicated channels while interactive traffic (NRT flows) use
shared channels. For fixed rate scheduling, the two 256 Kbps shared channels
are used. The DSCH code tree allows scheduling of a single user at high bit
rate or several lower bit rate users through code multiplexing.
For rate adaptive scheduling one of the two 256 Kbps branch can be set
aside to provide one 128 Kbps, one 64 Kbps and two 32 Kbps channels. In
this way, these schemes are using the same amount of code tree resources and
can be compared on a fair basis.
69
Fig. 3 through Fig. 9 depict the achieved performance for the interactive
service class. Looking jointly at user blocking, user satisfaction and system
throughput in KbpslMHz/cell for packet users, rate adaptation performs much
better than fixed rate scheduling across all traffic loads. Rate adaptation
achieves more than 90% user satisfaction and less than 1% blocking even at
high loads. Without rate adaptation the users satisfaction degrades to 60%
even if call blocking remains below 1%.
2.0
1.5
1.0
V
Without Rate Adaptation
w".~.~'.~
0.5
\
.. 0.0 ~_~
,,::;;;;;;;~J.;;;;;;;;;:~r.E::=;:::----'==2.~_-JI
I
0.2 O.~ 0.4 0.5
Arrival Rate I Second
Figure 3: % users blocked vs. arrival rate [s-l] of 256 Kbps Interactive Service
110
100
90
80
70
60
50
40r---------~----------~----------~
0.2 0.3 0.4 0.5
Figure 4: % users satisfied vs. arrival rate [s-I] of 256 Kbps Interactive Service
70
30
5r---------~----------~--------~
0.2 0.3 0.4 0.5
Arrival Rate! Second
Fig. 6 and Fig. 7 depict the average user throughput in Kbps during the
entire session and the normalized Service PDU (SPDU) transmission delay in
seclKbytes. SPDU delay includes queuing, transmission and retransmission
delays.
200
100
With Rate Adaptation
\
50r---------~----------~----------~
.17
.16
.15
.14
.13
.12
.11
.10
.09
.06
.07
.06 \ - - - - - - ; - - - - - - - - , - - - - - ,
0.2 0.3 0.4 0.5
Arrival Rate I Second
1.0
With Rate Adaptation
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
RT 4 Erlangs
0.1
0.0
36 38 40 42 44
BS Tx Power [dBm[
1.0
0.8
0.7
0.6
Without Rate Adaptation
\
0.5
0.4
0.:3
0.2
RT 4 Erlangs
0.1
NRT 0.4 Arrivalslsec
O.O\-.:w.....c=::!:::.----r--------l
10 20 :30
NRT Users Traffic Channel Tx Power [dBm[
Figure 9: CDF of traffic channel power [dBm) ofNRT Interactive class users
73
6
RT Load 4 Erlangs
4
Without Rate Adaptation
-1r----------.----------.---------~
0.2 0.3 0.4 0.5
Arrival Rate I Second of Interactive Users
100
90r--:::r--t--......._
80
70
RT Load 4 Erlangs
60r----------.----------.---------~
0.2 0.3 0.4 0.5
Arrival Rate I Second of Interactive Users
5. CONCLUSIONS
Results confirm that high rate users in bad channel conditions cause
increased interference and degrade performance. The introduction of rate
adaptation scheduling for NR T users leads to efficient use of radio resources
and better system stability. By giving priority to the RT conversational class,
only the remaining radio resources are allocated to NRT interactive class data
users. Rate adaptation scheduling results in better QoS for both service classes
and can accommodate more users by allowing bad channel NRT users to
transmit at lower rates. Combining rate adaptation scheduling for delay
tolerant services, prioritized CAC and power control is a promising path to
improve system performance and provide higher capacity for WCDMA
UMTS networks.
REFERENCES
[I] UMTS; QoS concept and architecture, 3GPP TS 23.107
[2] Gyung-Ho Hwang and Dong-Ho Cho, "Dynamic rate control based on interference and
transmission power in 3GPP WCDMA system", IEEE VTC'2000 Fall, vol.6, pp. 2926-
2931
[3] Takumi ITO, Seiichi Sampei and Norihiko Morinaga, "Adaptive transmission rate control
scheme for ABR services in the CBR and ABR services integrated DS/CDMA systems",
IEEE VTC'2000 Fall, vol. 5, pp. 2121-2125
[4] UMTS; Physical channels and mapping of transport channels onto physical channels
(FDD), 3GPP TS 25.21 1 V 3.4.0 Release 1999
[5] UMTS; "Selection procedures for the choice of radio transmission technologies of the
UMTS", (UMTS 30.03 V 3.2.0)
[6] A. lera, S. Marano, and A. Molinaro, "Call level and burst level priorities for effective
management of multimedia services in UMTS", IEEE INFOCOM '96, vol. 3, pp. 1363-
1370
Capacity And CII Performance Of Different Cell
Clusters In A Cellular Network
Abstract: Capacity and interference performance are among the most important issues in
the cellular frequency planning process. The main objective is to reach a
tradeoff between the quality and the offered traffic. In this paper, we study and
compare the CIR and spectral efficiency of different reuse patterns. We also
establish PDF expression of CII assuming one interferer in the serving cell.
The spectral efficiency and trunking efficiency are expressed analytically
versus cluster size, number of sectors and reuse distance. In particular, reuse
partitioning sub-cells sizes are optimized in order to maximize the traffic
capacity parameterized by spectral efficiency; and it is shown that more than
one partition wiI1 decrease offered traffic per area unit, per Hz.
Key words: Cellular Reuse Patterns, Interference, Quality Of Service, Capacity, Offered
Traffic, Spectral Efficiency, Trunking Efficiency, eIR Distribution Function,
Cluster Size, Reuse Distance, Reuse Partitioning (RP).
1. INTRODUCTION
Because of the limited available frequency bandwidth, cellular radio networks
adopt the frequency reuse concept to reuse the same frequency at different locations.
A large reuse distance can enhance the ell level by reducing the interference level
but offers a poor traffic capacity. One challenge for cell engineering is to reach a
tradeoff among channel quality, system capacity, and the costs of infrastructure and
user terminals. Previous works have been focused on frequency assignment
algorithms evaluation and comparison. In this paper, many already existing pattern
75
X. Lagrance and B. labbari (eds.),
Multiaccess, Mobility and Teletraffic for Wireless Communications, Volume 6, 75-86.
© 2002 Kluwer Academic Publishers.
76
types presenting particular frequency plans are studied and simulated considering
different reuse factors, antenna directivity and sectorisation types. We also consider
some particular cluster plans such as reuse partitioning (RP) and Fractional
Frequency Reuse.
The remaining of this paper is organized as follows: The next section compares
these clusters in terms of eIR performance on the basis of histograms plotted for the
same clusters families. An analytical PDF expression of ell is given in Section 3
assuming one interferer in the serving cell. In Section 4, the spectral efficiency and
the trunking efficiency are both calculated and expressed analytically. SeGtion 5
presents some cluster types examples and especially RP. We exhibit the best cluster
type that realizes the compromise between quality performance and traffic capacity.
Finally, our conclusions are summarized in Section 6.
"'~
(J) 80
," ,
,, ' . lOOOO< : K=4
. K=7 simple
::> ________ . : K=7 trapezoidal plan
u
Q) 70 ++++ : K=9
-5
.9 . K=12
2-
Q)
60
>
"m
~ 50
#'
.!: 40
"'
Q)
.c
"'
Q)
E
30
'0
Q)
0)
20
1ll
(J)
~
10
Q)
0..
0
0 10 20 50 60
In fact, the right side of each curve is unlimited whereas the one on the left has a
smaller boundary since we have calculated and represented C/I only within the cell
borders. Concerning the arrangement of the curves, we notice that, for the same
antenna type, the greater the cluster size K, the more the curve position is towards
the higher C/I values. So, C/I increases with cluster size. On the other hand, the
trapezoidal cluster presents the highest CIR values compared to all other patterns
because of its special configuration plotted in Figure 3 presenting only two potential
interferers versus six for the other cluster types. Only horizontal expansion is
possible in this latter plan; for this reason it is suitable for highways and coastal
areas [1].
6 effective interferers
Figure 4 depicts the sectorisation effect through the histograms of clusters with a
size equal to 3. We observe that the more the sectors of a site, the better the quality
for the same reuse factor. In fact, the 6-sectored cluster is more performing than the
tri-sectored ones that also present better CII than the simple omnidirectional cluster.
78
This is, mainly, due to the fact that sectorisation reduces the number of effective
interferers to 2 in the case of tri-sectored patterns and I for 6-sectored ones. Here,
the cyclic and alternate frequency assignments have comparable performance.
~ 50
#
.~
(f)
40
(])
.<::
(f)
(]) 30
E
'0
(])
0\
20
~ill
~
10
(])
D-
O
0 10 20 30 40 50 60
cn in dB
Figure 4. C/I histograms of clusters with a size 3 (Simulation).
Table I presents numeric results for some clusters. We can conclude that the
cyclic channel assignment provides a better ell quality than the alternate assignment
with a value of more than 4 dB for the cluster 4112. This difference is due to the fact
that, for the alternate assignment case, different cells antenna beams are interfering
by 60° each from the other inducing more interference than the cyclic assignment
cluster where co-channel cell beams are pointing to the back of each other for which
secondary lobes imply less interference.
Another crucial factor for ell quality is the number of effective interferers. It
appears in the case of two 4/12 clusters that have the same reuse distance but the
fIrst has six potential interferers and the second has only four, which increases ell
by more than 2.5 dB than the fIrst model.
79
3. C/I ANALYTICAL DISTRIBUTION MODEL
(2)
BTS j
(3)
r
The CDF F is then given by
~ ::: .,t.~.,.,
eoo 0.07 .
_ en probability density (PDF)
~004
.B ;
~ 0.03 :
0: :
0.02 :
0.Q1
15 20 w ~ ~ ~ ~ ~ ~
C/ljndB
E = EI'QIK-IC~B). (10)
BtottrR
By extracting the reuse distance D from its approximation T3K .R valid for
regular clusters, (10) becomes
3fp (K)
E= 8 2' (11)
BtotlrD
where fpo is the characteristic function defined by
!p.(n) = nE1,Qln-1CPB). (12)
Its variation is shown to be decreasing ; since the spectral efficiency E is
proportional to fp (K) , so it decreases with the increase of the cluster size K and is
inversely proportional to the reuse distance squared.
The trunking efficiency expression is
E -I(p')
TE = 100 I,QIK B = 1001; (K). (13)
QIK Q p.
CD
-0
eo
,5
i5 40
20
0
0.5
-3.5
Q)
(J) -4
co
~
~ -4.5
9-
Q)
u
c -5
Q)
Q5
~ -5.5
>-
u
c
Q)
'u -6
t
Q)
g> -6.5
:2
c
~ -7
-7.5
-8
2 3 4 5 6 7 8 9 10 11 12
Cluster size of the ring
Figure 9 We note the existence of negative values of the curve especially for Km
:$ 0.53 K where the ordinary cluster is better than RP. This performance reduction
J
reaches its maximum when Km = 0.24 K J and Ks = 0.5 Km = 0.12 K J with a 10-2
approximation where the capacity loss is 19.89%. However, the spectral efficiency
gain is maximum when Km = KJ and Ks = 0.5 Km = 0.5 K J where we have a 29.33
gain. The addition of more than one partition decreases spectral efficiency
performance. This result differs from results obtained in [4] where the bandwidth
increases indefinitely as long as we add partitions. This is due to the fact that for the
case of the addition of more than one partition, the bandwidth gain is more than
compensated by the trunking efficiency loss described in Sub-Section 5.2.
Percentage of spectral efficiency increase for "Reuse Partitioning" in relation to ordinary cluster
30
.s
Q)
II)
20
(II
Q)
g 10
(;'
c:
3 0
~
~
u
-10
'"a.
(f) -20
1
Figure 9. Spectral efficiency gain variation of RP cluster versus sub-cells relative dimensions.
However, the last cluster is better than the other one by about 3 dB due to its
bigger reuse distance. In fact, the tri-sectored cluster has a 7-sized cluster, which is
bigger than that of Fractional Frequency Reuse (5.333).
REFERENCES
Li-Chun Wang
Ching-Yu Liao
and Chung-J u Chang
Department of Communication Engineering,
National Chiao Tung University
Tel: +886-3-5712121 ext. 54511
Email: {lichun.cjchang}@cc.nctu.edu.tw.cyliao.cm86g@nctu.edu.tw
87
X. Lagrance and B. labbari (eds.),
Multiaccess, Mobility and Teletrafficfor Wireless Communications, Volume 6,87-102.
© 2002 Kluwer Academic Publishers.
88
1. Introduction
Soft handover is one of the most important merits of the code division
multiple access (CDMA) cellular system. For user terminals in the soft
handoff process, the original serving base station and the target base
station will maintain two communications links simultaneously over the
same bandwidth to guarantee a smooth transition without the chance
of dropping the ongoing call. Traditional soft handoff algorithms are
developed for homogeneous cell structures, i.e, the involved base stations
have the same cell size. In practice, however, cell coverage area of each
base station in the existing cellular network differs a lot from each other.
First, due to the coverage problem, a small micro cell may be installed
at the boundary of surrounding macro cell base stations. Second, for
increasing system capacity, a cluster of micro cells may be employed by
cell splitting or other techniques. Thus, a heterogeneous cellular network
will be naturally occurred as shown in Figure 1.
soft handover is occurred between two cells with different cell radius.
According to the EPA method, the handoff mobile terminal originally
in the macro cell will request the same amount of power to be allocated
from both macro cell base station and micro cell base station. Intuitively,
it is easy to see that the handoff users form macrocell will be very likely
to exhaust most of the power budget in the micro cell which usually
has a small amount of total transmission power, thereby decreasing the
regular user capacity of the microcell. However, to what extent the ratio
of cell radius between two serving base stations affects the effectiveness
of the power allocation method in the performance gain of downlink soft
handover is an open issue.
To overcome the power exhausting problem of the EPA method in
downlink soft handover within heterogeneous cellular networks, we pro-
pose a new quality balancing algorithm with power constraints for allo-
cating the base station transmission power to serve each user in the both
macrocell and microcell. We call the new power allocation algorithm
for soft handover the constrained unequal power allocation with (UPA)
method in this paper since the allocated power form the serving base
station will be different. The constrained UPA method is the modified
version of [4]. Unlike [4] suitable for only a homogeneous cell structure,
the proposed constrained UPA method is applicable for the heteroge-
neous cell structures. More importantly, we add a new criterion to limit
the maximum downlink transmission power allocated from base stations
to each handoff user. The new criterion is based upon the link budget
analysis to calculate the transmission power to achieve the required SIR
requirement at the cell boundary. According to the constrained UPA
method, all the allocated power from the base station for any handoff
request should be constrained below this limit. Our numerical results
will demonstrate that the concept of limiting maximum downlink trans-
mission power is very critical to avoid the power exhausting problem
occurred in the micro cell especially when the cell radius of the micro cell
is less than 50% of that of the macrocell.
The literature survey of the previous work related to downlink soft
handover for heterogeneous network can be classified into three cate-
gories. First, the CDMA downlink soft handover issue was first exam-
ined in [1], where the impact of downlink soft handover was discussed
for a homogeneous cell structure. In [1], it is mentioned that the EPA
method is effective for a mobile station during soft handover, but the
detailed algorithm and downlink capacity with soft handover is not ana-
lyzed. The effect of soft and softer handovers on the downlink capacity of
homogeneous CDMA was discussed in [5], but without addressing the is-
sue of power control. Secondly, as for the downlink power control issues,
many downlink power control algorithms for balancing link quality have
been developed in both centralized version and distributed version [6]
90
and [4]. The concept of quality balancing power control is to let all the
mobile stations in the cell maintain equal link quality. On the contrary,
the conventional method will allocate extra transmission power, result-
ing in extra interference. Few downlink power control works have been
published in the context of heterogeneous cellular networks. At last, for
the mixed cellular architecture, many research works have be published
to analyze the performance of a hotspot micro cell embedded within a
larger macro cell [7] and [10]. Few work except [11] has been published
to discuss the performance for the mixed cellular architecture as shown
in Figure 1. However, in [11], only the reverse link capacity without soft
handover in a CDMA cellular system with power control with mixed cell
sizes was analyzed. Thus, to our knowledge, the performance analysis
of downlink soft handoff with power control for heterogeneous cellular
networks is still an open research area.
As for the uplink performance analysis for heterogeneous CDMA net-
works, [12] investigated the interference issue when a mobile terminal
connecting the macrocell moves toward a mobile terminal connecting
the microcell. Although the reverse link is considered to be a limiting
factor for the CDMA system capacity, the forward-link performance is
becoming increasingly important due to the emergence of asymmetric
wireless data services.
In summary, this paper evaluates the capacity of a heterogeneous cel-
lular network with a hotspot micro cell adjacent to a larger macrocell. We
consider the effects of path loss, shadowing, multiple access interference,
downlink power allocation, soft handover in our performance evaluation.
Furthermore, we present a new quality balancing downlink power allo-
cation method with maximum power constraint, the constrained UPA
method, which can solve the power exhausting problem in the micro cell
when executing soft handover. Using the constrained UPA method, a
smaller micro cell can be employed without suffering the power exhaust-
ing problem, thereby increasing the total system capacity.
The remaining parts of this paper are organized as follows. Focusing
on a simplified cell model with a single micro cell and a single macrocell,
Section 2 first discusses the signal model, downlink power control, and
soft handover gain, and then presents an analytical model for comput-
ing the downlink user capacity with soft handover in the heterogeneous
cellular network. The numerical results are shown in Section 3. Section
4 will give the concluding remarks.
91
2. System Model
2.1 Signal Model
Consider a simplified heterogeneous cellular model with a single mi-
crocell adjacent to a macrocell as shown in Figure 2. Let RM and R ft ,
respectively denote the radii of the macro cell M and the micro cell /L;
r M and r ft are the distance for the user terminal at the point H to the
macro cell M and that to the micro cell /L, respectively.
,
"-
\
\
Il I
I
/
/
qi,j . Li,j . G
r(qi,j) = (P't _ q .. ) .L·· + L P k . L k · +7] ;::: 'Yreq, (1)
1,,) 't,) ,) 0
k,kfi
N
where G is the processing gain, Pi = L qi,j is the total downlink trans-
j=1
mission power for N users in cell i; Li,j is the link gain from cell i to
mobile j, 7]0 is the background noise, and 'Yreq the required Eb/No. We
include the effect of both path loss and shadowing in the link gain Li,j.
That is,
Li ,J' = L'1,,). X lO~dlO (2)
In (2), L~,j follows a two-slope path loss model as in [7].
L'
A
'"
{ dZ~-:/3
..
,if di,j > > Zi
(3)
i,j = da. (1 + (~){3) , = (~'~{3Si)
'b,) Zi
dO. Zz
, J'f di,j «Zi
where a and f3 are the path loss exponents, di,j is the distance from
mobile station j to the base station i, Zi is the break point in cell i, and
A is a constant. The standard deviation of the shadowing ~i in (2) is
92
,di,j ::; Zi
(4)
,di,j ~ Zi
(5)
where hi is the antenna height for base station i, h ms the mobile antenna
height, and ,\ the wavelength. We define the cell boundary as the point
at which mobile station j receives the same power from both adjacent
cells M and f-l [9]. Then at the cell boundary, we have
(6)
For simplicity, we only consider the effect of path loss in (6). Then,
combining (3) and (6)
P L' + (!lM..)f3)
Rcx (1 R h
~M = ---1!:2i... = M ZM ex (-.!:!... t+f3 x (---1!:....)f3 (7)
PI' LM,j R~(l + (~)f3) RI' hM
Note that (7) is only valid when the micro cell radius is higher than the
break point distance. From (I), where the noise is neglected and only
micro-cell interference is considered, we have
(8)
where
(d-CXJ-L (1 + ~ )-f3J-L)
D. - I' ZJ-L (9)
) - (dA:tM (1 + ~ )-f3M)
To make macro cell users have the required Eb/ No, the maximum al-
locating transmission power qM can be obtained by substituting the
maximum total transmission power PM and PI' in (8). Then
(10)
where D j is given in (9). Note that the total transmission power of the
base station is varied depending on the summation of power allocated
for each user. The hat in qM indicates that this power level allocated
from the base station is for the user at the cell boundary. From (7) and
93
(10), the downlink maximum allocating power for the micro cell can be
obtained as
~ ~
L'M,j ( )
q/1 = qM' ~ 11
/1,)
where L~,j and L~,j are given in (3). In this paper, we adopt maximum
ratio combining for the forward link soft handover. Thus, based on [15],
the optimal received Eb/No for mobile station j during soft handover is
given by
r( qM,j, q/1,j) = r( qM,j ) + r( q/1,j) , (12)
where r( qM,j, q/1,j) denote the Eb/No after the maximum ratio com-
bining for macrocell transmitting at the power level qM,j and micro cell
transmitting at q/1,j, respectively; r( qM,j) and r( q/1,j) is the E b/ No re-
ceived from the macrocell base station and that from the micro cell base
station before combining, respectively.
For the soft handover case, according to (12), the soft handover gain
Csojt can be obtained by
Note that only the mobile terminals j E Ii - Ti will involve the quality
balancing power control, whereas the mobiles j E Ti will use the Pi. The
allocated procedures for each mobile station j in cell i is given by
_
qi,j = (Pi -
Q)
i
¢i,j
~,/, .. (18)
L..J '/",J
JEIi
jetTi
(21)
Set the stop criterion as I1'M - 1'111 c < 10- 3 . The power control
algorithm is described as follows:
• Step 0: [Initialization]
Set cO > > 10-3 , iteration w = O.
For all cell i E M, p,: set Q? = 0, TP = 0, N¥(i) = 0, ~o = N2(i)'(k
• Step 1: [Power allocation for mobiles not in T i ]
Obtain iir,j and the corresponding 1't' by (21) and (18), respectively.
• Step 2: [Decision of power control procedure]
IF cW > 10- 3 , CONTINUE Step 3.
ELSE GOTO Step 4.
(23)
Given that number of soft handover users in macrocell M and micro cell
J.L (NJ.2°, N~ho), we can calculate outage probability for macro cell M
from (8)
(N M-N;';O) N;;o N~ho
Note that the first term in (24) can be used to incorporate the effect
of constrained downlink power allocation, and the second term and the
third term incorporate the effect of UPA or EPA. Let
(NM-N!!1°)
YM = L Dj · 10(!:I"-!:M)/10 (25)
j
and
where Q(x) = Jxoo ~e-t2 /2dt. Note that since YM is a sum of independent
log-normal random variables, it can be approximated by another log-
normal random variable YM with mean my and standard deviation (Jy
by using the techniques in [16]. The outage calculation is validate both
for micro cell and macro cell in the forward link.
3.2 Discussion
In this paper, we study the performance of the soft handover in CDMA
cellular systems with mixed cell sizes. At first, we analyze the perfor-
mance of the received Ebj No for a mobile station in the location H as
shown in Figure 2. Figure 3 shows the EbjNo performance in different
cell radius ratio p supporting soft handover mobile H by EPA and UPA
methods, respectively. In the EPA case, according to (13), executing
soft handover can improve received signal quality in homogeneous cellu-
lar environment, i.e., p = 1.0. However, the received signal quality will
be deteriorated for mobile station h moving from microcell to macrocell,
denoted by f.L ---> M in the figure, as soon as the cell radius ratio is get-
ting smaller. More importantly, we find that EPA method will result in
the power exhausting problem when the cell radius ratio is smaller than
0.3, for which the micro cell consumes too much power in supporting
soft handover mobile stations moving from macrocell, thereby leaving
no power budget to support its own users. On the contrary, the UPA
method, based on (14), the EbjNo can be maintained at the satisfactory
level for p = 0.1 rv 1.0. Thus it is demonstrated that the UPA method
can avoid the power exhausting problem in micro cell for smaller cell ra-
dius ratio.
18
16 -e- M-Il , EPA
-(3- /l-M, EPA
14 -e- M-I', UPA
-El-Il - M. UPA
12
co 10
hard
soft handoff
~
0 8 andoff
z
:0
w 6
.-.-.-.~.-.-.-.
cr--
4 ~~
2
-
---
microcell
0 macrocell
no handoff
-2 0
0.2 0.4 0.6 0.8
Cell radius ratio (p)
Figure 3. Eb/No performance in different cell radius ratio for mobile station h
without handover and with soft handover by equall power allocation and unequall
power allocation methods, respectively. J.L ~ M means a user moves from a microcell
f.1 to a macrocell M, and M -> f.1 means a user moves from a macro cell to a microcell.
p = 1.0, EPA
~
:cI1l
.c 10 ·1
e
D-
Q)
0>
.:9
:::J
0
10 ·2
20 25 30 35 40 20 25 30 35 40
Capacity (number of mobile stations) Capacity (number of mobile stations)
Figure 4- Outage probability of microcell and macro cell using conventional quality
balancing power control method with equal power allocation (EPA) soft handover
algorithm for p = 0.7, 1.0, where p is the cell radius ratio of micro cell and macrocell.
macrocell
~ 36 35
III
'":::>
'0 32 30
iii
.0
.s
E
28 25
o No power constraint
o Power constraint
~
u
~ 24 20
to ONo power constraint
U o Power constraint
Figure 5. Capacity of (a) equal power allocation (EPA) and (b) unequal power
allocation (UPA) with soft handover against the ratio of radius of the microcell to
that of the macrocell p.
nearly 10 times less than that by simulation. From the figure, one can
see that the errors in capacity approximation are from zero to two users,
which is tolerable for capacity calculation. Thus, this proves the accu-
racy and efficiency of using this analytical approximation in calculating
the user capacity for the heterogeneous cellular systems.
Figure 5a shows the capacity of the EPA method with soft handoff
against the cell radius ratio p, in terms of outage probability equal to
0.05. It is observed that the power exhausting problem occurred in the
micro cell for p < 0.7 without power constraint and p < 0.5 with power
constraint, respectively. One can see that the smaller the value of p,
100
65
60
~ 55
~ 60
'0 50
2i
I.??
45 UPA
55
40
[ 35
u" 50 L:'-c--'--c:'---'--~~~~~-.J
30
0 0.1 0.3 0.5 0.7 0.9 1 0 0.1 0.3 0.5 0.7 0.9 1
Cell radius ratio ( P ) Cell radius ratio (p )
(a) (b)
Figure 6. Total capacity comparison of the equal power allocation (EPA) and
the unequal power allocation (UPA) during soft handover with and without power
constraint.
the higher the macro cell capacity. The increase of macro cell capacity
is mainly because of the decreasing interference from the microcell and
using up all the power budget of the micro cell base station. Hence,
although constraining the maximum power can help solving the power
exhausting problem in the micro cell, the improvement is not significant.
Figure 5b demonstrates the capacity of the UPA method with soft han-
dover against the cell radius ratio in terms of 0.5 of outage probability.
Unlike the EPA method, the UPA can maintain a good capacity for
both micro cell and macrocell from p = 0.5 '" 1. The power exhausting
problem will not occur even with p = 0.1 although with slightly capacity
degradation in microcell. It is also noted that the power constraint can
improve the capacity, especially when the p is small. For p = 0.1 the
capacity for the constrained UPA method increases microcell capacity
about 30%.
Figure 6 shows the total capacity of EPA and UPA methods. The
total capacity here is the summation of the macrocell capacity and the
micro cell capacity in Figure 5. The purpose of plotting Figure 6 is
to calculate the capacity in the case of one macro cell and a cluster of
microcells. We assume the coverage is equal to two macrocell coverage
area. Based on the area's ratio (i? between macro cell and microcell, the
estimated capacity of one macrocell and a cluster of multiple micro cells
can be calculated as (NM + N", x (i)2), where NM and N", is the user
capacity of a macro cell and a micro cell, respectively. As shown in Figure
7, the concept of constraining the maximum transmission power during
hand off is helpful in avoiding the power exhausting problem only for
p> 0.5. As the star marks in the figure, the EPA method suffers from
the power exhausting problem for p = 0.3 and 0.5, where the capacity of
multiple micro cells is sacrificed for only a macro cell capacity. The UPA
method can provide high user capacity regardless of power constraint.
For p = 0.3, it is easily to see that the UPA method can enhance total
101
system capacity eight times higher than the EPA method. Hence, it is
concluded that the UPA method is an inevitable technique in solving the
power exhausting problem during the soft handoff in a CDMA network
with mixed cell size.
400
.::I no power constraint with EPA *
*
3~5
350 • _ consl",int wit/1 EPA
319 ~ no power constrainl wilh UPA
~ 300 • power oonstraint with UPA
Q)
en
::J
'0 250
Q;
.c 200
E
::J
E- 150
~
ro
u 100 84 91 91 94
c..
ro 60 62 6062
u
50
~ 0
0.5 0.7 1.0
Cell radius ratio (p)
Figure 7. Approximate capacity of one macro cell and multiple microcells for power
control algorithms with and without power constraint, combined with EPA and UPA
methods for soft handover power allocation, where there are 2 microcell for p = 0.7,
4 microcell for p = 0.5, and 333 microcells for p = 0.3, respectively.
4. Concluding Remarks
This paper studies the downlink user capacity of a heterogeneous
CDMA cellular system with soft handover. We consider the scenario
that a hotspot microcell is adjacent to a larger macrocell. We ob-
serve the phenomenon of the power exhausting problem happened in
soft handoff between micro cells and macrocell. To quantize the impact
of this problem, we present an analytic approximation method for com-
puting the downlink user capacity with soft handover in heterogeneous
cellular structures. We further propose an improved quality balancing
power control method, the constrained unequal power allocation (UPA)
method to protect micro cell base station from being used up the trans-
mission power by the hand off mobile terminals in macrocell.
Our simulation results demonstrate that the proposed constrained
UPA technique with soft handover can enhance the total system ca-
pacity eight times higher than the conventional equal power allocation
(EPA) method in a heterogeneous CDMA network with the cell radius
ratio between micro cell and macrocell equal to 0.3. Future work in this
area include to extend the analytical capacity estimation technique to
multiple clusters of micro cells with multiple macrocells, and develop an
optimal downlink power allocation algorithms with soft handover for the
CDMA network with heterogeneous cell structures.
102
References
[3] Holma H., and Toskala A., "WCDMA for UMTS: radio access for third gener-
ation mobile communications," John Wiley and Sons, ltd., pp. 208-210, 2000.
[4] Kim D., "A simple algorithm for adjusting cell-site transmitter power in CDMA
cellular systems," IEEE Trans. on Veh. Technol., vol. 48, no. 4, pp.1092-1098,
July 1999.
[5] Lee C. C. and Steele R., "Effect of soft and softer handoffs on CDMA system
capacity," IEEE Trans. on Veh. Technol., vol. 47, no. 3, pp. 830-841, Aug. 1998.
[6] Grandhi S. A., Zander J., and Yates R. D., "Constrained power control," Wire-
less Personal Commun., vol. 1, no. 4, pp. 257-270, 1995.
[7] Erceg V., Ghassemzadeh S., Taylor M., Li D., and Schilling D. L., "Ur-
ban/suburban out-of-sight propagation modeling," IEEE Commun. Mag., pp.
56-61, June 1992.
[8] Min S., and Bertoni H. L., "Effect of path loss model on CDMA system design
for highway microcells," IEEE VTC98, pp. 1009-1013, 1998.
[9] Shapira J., "Microcell engineering in CDMA cellular networks," IEEE Trans.
Veh. Technol., vol. 43, no. 4, pp. 817-825, Nov. 1994.
[10] Shalinee K., Greenstein L., Poor H. V., "Capacity tradeoffs between macro-
cell and microcell in a CDMA sysem: exact and approximate analyses," IEEE
Vehicular Technology Conference, VTC'Ol Fall, Atlantic City, pp. 1172-1176,
October, 2001.
[11] Jeon H. G., Shin S. M., Hwang T., and Kang C. E., "Reverse link capacity
analysis of a CDMA cellular system with mixed cell sizes," IEEE Trans. on
Veh. Technol., vol. 49, no. 6, pp. 2158-2163, Nov. 2000.
[12] Lee D. D., Kim D. H., Chung C. Y., Kim H. G., and Whang K. C., "Other-cell
interference with power in macro/microcell CDMA networks," IEEE Vehicular
Technology Conference, pp. 88-92, 1996.
[13] Kim J. Y., Stuber G. L., Akyildiz 1. F., "Macro diversity power control in hier-
archical CDMA cellular systems," IEEE J. Select. Areas Commun., vol. 19, no.
2, pp. 266 V276, Feb 2001.
[14] Andersin M., Rosberg Z., and Zander J., "Gradual removals in cellular PCS
with constrained power control and noise," ACMjBaltzer Wireless Networks J.,
vol. 2, no. 1, pp. 27-43, 1996.
[15] 3GPP technical Specification 25.942, RF System Scenarios, page 26, Dec. 1999.
[16] Schwartz S. and Yeh Y. S., "On the distribution function and moments of power
sums with log-normal components," Bell System Tech. Journal, vol. 61, pp.
1441-1462, Sept. 1982.
Packet service in UMTS: effects of the
radio interface parameters on the
performance of the downlink shared
channel
Flamini Borgonovo, Antonio Capone, Matteo Cesana, Luigi Fratta
DEI, Politecnico di Milano
borgonov,capone,cesana,fratta@elet.polimi.it
Abstract
The UMTS W-CDMA radio interface is characterized by great
flexibility and a variety of different physical and logical channel
types. For example, on the downlink, the DCH offers circuit switch-
ing, the FACH uses packet switching and the DSCH uses packet
switching with closed loop power control. Furthermore, several user
rates and protections are possible, by choosing suitable parameters,
such as spreading factors, code rates and ARQ schemes. In this
paper we present the results, obtained by a detailed simulation,
about the effect of several parameters and system alternatives on
the capacity of the downlink segment of the W-CDMA interface
with packet service. In particular, we investigate the effect of the
spreading factor and the code rate on the DSCH capacity and delay-
throughput performance.
1 Introd uction
The Universal Mobile Telecommunications System (UMTS) [1, 2] is the third
generation mobile communication system developed by ETSI, the European
Telecommunications Standard Institute, which will allow the use of a new fre-
quency spectrum and is expected to extend the present GSM service to include
multimedia.
In UMTS, users will be provided with data rates up to: 144 kb/s, in macro-
cellular environments, 384 kb/s, in micro-cellular environments, and up to 2 Mb/s
in indoor or pico-cellular environments. Due to the effort of the standardization
bodies, the radio interface is characterized by great flexibility and a variety of
different physical and logical channel types. For instance, several user rates
and protections are possible, by choosing suitable parameters, such as spreading
factors, code rates and ARQ (Automatic Repeat request) schemes.
103
X. Lagrance and B. Jabbari (eds.),
Multiaccess, Mobility and Teletraffic for Wireless Communications, Volume 6, 103-114.
© 2002 Kluwer Academic Publishers.
104
Among the new services offered by UMTS, the packet data service is prob-
ably one of the most critical from the system parameters setting point of view
mainly because of the characteristics of the code division access scheme adopted
at the radio interface. Up to date no study that thoroughly investigates the
effects of the different possibilities on UMTS data service performance has yet
appeared.
In the downlink, three different transport channel types are available for data
packets transmission, namely the DCH Dedicated Channel, the DSCH Downlink
Shared Channel and the FACH Forward Access Channel.
The DCHs are assigned to single users through set-up and tear down pro-
cedures and are subject to closed loop power control that, if used for circuit
service such as voice, stabilizes the BER (bit error rate) and optimizes CDMA
performance.
The DSCH is a common channel on which several users can be time mul-
tiplexed. No set-up and tear down procedures are required and the physical
channel on which the DSCH is mapped does not carry power control signaling.
However, since the closed loop power control is still required, users that are al-
lowed to access DSCH services must have an associated active DCH. The DCH,
if not already active due to another transport service, must be activated just to
allow the access to the DSCH and to carry physical layer signaling only.
The FACH is shared by several users to transmit short bursts of data, but,
unlike DSCH, no closed-loop power control is exerted and no DCH must be
activated to access this channel.
For each one of the above channels, different combinations of spreading factor
and code rate can provide the bandwidth and the protection required for different
services and environments. However, it is not altogether clear which combination
is the best.
Well known results for real-time circuit traffic show that CDMA with closed-
loop power control can be very effective in spectrum exploitation [3]. Its effi-
ciency can be further enhanced by using powerful codes and FEe codes have
been proved to be more effective than spreading codes [4].
With packet service, the effect of direct sequence spreading, FEC codes and
closed loop power control is not easily predictable and the optimal combination
of codes and spreading factors may be different from circuit service. In fact,
data traffic is bursty in nature, and, depending on the number of interfering
channels and their power levels, errors can be more efficiently obviated by ARQ
techniques than by forward error correcting codes [5, 6]. For the same reason,
the protection obtained with high spreading factors is questionable.
To understand the roles that the many parameters and system features have
on the overall capacity with packet service, we have implemented a detailed
simulator of the UMTS downlink.
In Section 2 we present the system model adopted for simulations and in
Section 3 we discuss the results obtained. Section 4 includes some final remarks
and concludes the paper.
105
2 Simulator description
The simulator reproduces a system composed of 49 hexagonal cells that lay on a
torus surface to avoid border effects. The base stations (BS) are located at the
center of each cell and irradiate with omni-directional antennas with unit gain.
I.E-Ol t---"''''-~*-------I
I.E-01 t-----t--H---1.----j
I.E-02 +-----+1r1-~-----l
n: n:
~1.E-03 t-------1M-\-----''<----I w
...I
III1.E-02
I.E'()4
I.E-05 +-~-_~~---'.,u..~-~.>...,_-l
-604-2024 10 12 -6 -4 ·2 0 2 4 6 8 10 12 14 16
Eb/No [db] SIR [db]
Before transmission, the physical layer adds the redundancy bits according to
the coding scheme adopted. Several coding schemes are supported by UMTS.
Our simulator adds the parity bits required by Convolutional Codes, with
256 states, Constraint Length K = 9 and optimal puncturing, whose Bit Error
Rate (BER), obtained through link level simulations [10], is shown in Figure
1. In particular we have considered code rates, spreading factors and block
sizes such that the bits introduced by rate matching are very few and add an
overhead, without increasing error protection. To avoid throughput differences
due to different mappings of bits from packets to blocks we have used almost
the same transport-blocks size for all codes and spreading factors. A block
length of about 750 bits has been proved optimal with respect to the maximum
throughput.
c
I
= ~~--~--~
O'.lintra + linter + PN
(1)
Eb 1 C
-No = -2R xSFx-
I'
(2)
where R is the coding rate and SF the spreading factor. The term ~ comes
from the fact that since QPSK modulation scheme is adopted on the downlink of
UMTS-FDD, each information symbol is composed of 2 bits. The ratio Ebl No
represents, in fact, the entry in Figure 1. From the curves shown in this figure,
BLock Error Rate (BLER) curves have been obtained as BLER = 1- (1- BER)1 ,
1 being the transmission block length. For each transmission, the normalized bit
energy is used to derive the BLER, and the correctness of the transmission is
decided testing the value of a normalized random variable against BLER.
Otherwise specified, in the following we will refer to the SIR after despread-
ing, which is defined as SIR=SF x C/I. Figure 2 reports the BLER of blocks of
750 bits versus SIR.
Our simulator does not implement an explicit ARQ procedure. Instead, at
the end of the transmission, the transmitted block is kept in the transmitting
queue unless no error occurs. After 10 failed transmissions the block is dropped
and the user is declared in outage.
No
Pj "'" Pmax ) - - - - - '
hes
I Ni :=Ni+ll
+
Pick Fe_timer,
in
[O:N,]
a stable behavior, it's necessary limit the average interference level. We have
proposed and implemented in our simulator a flow control mechanism based on
the well known back-off (BO) mechanism which dynamically adjusts the load on
the active channels, and therefore the average interference generated, according
to traffic and propagation conditions.
The BO mechanism is based on a feedback provided by mobile terminals.
The basic idea is to reduce the transmission rate on the channel when one or
more consecutive transmissions fails. The mechanism is triggered only when
a transmission performed at the maximum powers. More in details, a flow-
control timer is started so that transmissions are inhibited for a random number
of frames (10 ms long) uniformly distributed in the interval (l,n), where n is
the number of consecutive wrong transmissions performed at maximum power.
In such a way, we control the traffic on the channels (G), and consequently
we limit the mean interference level. In order to let the information on the
transmission result at the transmitting end available, we adopt one of the FBI
(Feedback Information Bits) bits defined in the uplink transport block format
of the dedicated channel (DCH) associated to the DSCH [12]. The flow chart of
the BO mechanism is reported in Figures 3 and 4.
3 Simulation results
The complexity of the overall system and the interaction among system pa-
rameters and performance variables do not allow a single and straightforward
discussion of the system behavior. In our investigation we have been forced to
focus the study into several sub-problems and to take simplifying assumptions.
109
2000 1200
SF=4 SF=4
! iii'
1000
..
:a
1500
R=2/3
BOO
>-
CO ~
..
'ii :::I
"C 1000 C. 600
..c
....
CI
CO
.
CI
:::I
0 400
SF=4
-
c R=1
o
:s 0.1 .
.Il 0.1
/ R-3/4
,!g a: ./
c
o
g
UJ
~
R=3/4
~:::I ""U
0.01 .2 0.01
111 _____ R-1I2
~ SF=4
behavior is observed with the code R = 2/3, though the G limit is further in-
creased to 0.97. The delay curves for the two last cases almost perfectly overlap,
showing the ability of the BO to keep the system very close to capacity. How-
ever, although the reduced SIR target reduces the saturation fraction and the
BLER, the obtained BLER is still much higher than what predicted by Figure 2.
The reason is that, the power control mechanism is able to track the SIR at the
target value only with slow varying interference, as in the case of circuit service,
but fails its goal with packet service where traffic and interference are bursty.
We have measured SIR standard deviation values in the range 3.7 - 4.3 dB 2.
With more powerful codes (R= 1/2), G = 1 can be reached and, because of
the further reduced SIR target, the saturation fraction and the BLER is further
reduced. However, the benefits of the reduced number of retransmissions, do not
compensate the loss of throughput due to the increased code redundancy and
the maximum observed throughput is remarkably smaller than that obtained
with R= 2/3.
The system parameters configuration with R= 1/2 is very effective in fighting
interference, since it allows transmissions with lower power levels. It's quite clear
from figure 5 that the redundancy introduced by the code R= 1/2 limits the
capacity of the system, when using a single PDSCH. Under these hypothesis we
have studied the performance of the DSCH service when using multiple SF= 8
in a single BS, and a FEC code with R= 1/2.
The performance is significantly improved, as observed in Figure 9, by using
3 and 4 PDSCHs with a SIR target equal to 4 dB, which in this case has shown to
be optimal. The BO mechanism intervenes with 4 PDSCH limiting the maximum
G to 0.855. With 5 PDSCHs, the increase in the interference prevails, G is limited
to 0.622, and a small instability effect is present despite the BO. The case of 4
channels provides the maximum throughput (1240 kb/s) among those examined.
4th
SF=8 ,eh b.D.
R=1/2 b.o
'"f!
3
.. 500
o+---~-~-~-~--~-~
o 200 400 600 800 1000 1200
Throughput (kb/s)
DCHs which affects the performance of the DSCHs. Note that limiting to N
the number of active DCH has from one side the beneficial effect of reducing
the interference, but from the other it reduces the efficiency of multiplexing and
might penalize the achieved throughput in the case of sources with low speed.
Therefore an optimal choice of N exists for any given traffic characteristics. In
our case, which assumes the basic traffic model described in section 2, N = 10
is an optimal choice since we did not observe any multiplexing inefficiency.
4 Conclusions
In this paper we have investigated the performance of the UMTS radio interface,
with packets service, mainly evaluating the maximum throughput achievable on
the DSCH with different physical channels configurations, traffic dynamics and
power control mechanisms.
The results show that, when the packet service can use only one single phys-
ical channel, the maximum throughput is attained with the smallest available
spreading factor, SF= 4, and a light code, R = 2/3. The other cases character-
ized by a higher channel protection (lower R ) present a lower throughput since
the loss due to the added overhead is not compensated by the reduced BLER.
If the use of multiple physical channels is allowed, the maximum throughput
is attained by using up to four channels with SF= 8 and R = 1/2, despite
the new intra-cell interference introduced. This is mainly due to the improved
efficiency of the closed-loop power control that, taking advantage of the longer
transmission time and of the reduced interference burstiness, better tracks the
SIR target. This confirms the common belief that CDMA characteristics are
better exploited by circuit services with constant interference and, therefore, by
using many small channels rather than a single big one.
Further investigations, still in progress, show that the conclusions drawn
113
2000
N-10
SF=8
R=1/2
r-- 4 channels
..,
N-20
en 1500
.§.
~
~
., 1000
.,E'"
~ 500
__ I.)
o
o 500 1000 1500
Throughput (kb/s)
here somehow scale to other scenarios, such as micro-cells and/or corner fed
antennas.
References
[1] A. Samukic, UMTS universal mobile telecommunications system: development of
standards for the third generation, IEEE 'Transactions on Vehicular Technology,
vol. 47, no. 4, Nov. 1998, pp. 1099-1104.
[2] K.W. Richardson, UMTS overview, Electronics & Communication Engineering
Journal, vol. 12, no. 3, June 2000, pp. 93-100.
[3] A.M. Viterbi, A.J. Viterbi, Erlang capacity of a power controlled CDMA system,
IEEE Journal on Selected Areas in Communications, vol. 11, no. 6, Aug. 1993,
pp. 892-900.
[4] J. Y. N. Hui, Throughput analysis for Code Division Multiple Access of the Spread
Spectrum Channel, IEEE Journal on Selected Areas in Communications, vol. 2,
no. 4, July 1984.
[5] R.J. McEliece, W.E. Stark, Channels with block interference, IEEE 'Trans. on
Information Theory, Vol. 30, No.1, January 1984.
[6] F. Borgonovo, A. Capone, L. Fratta, Retransmissions Versus FEC Plus Interleav-
ing for Real·Time Applications: A Comparison Between CDPA and MC-TDMA
Cellular Systems, IEEE Journal on Selected Areas in Communications, vol. 17,
no. 11, Nov. 1999.
[7] UMTS 30.03, Annex B: Test environments and deployment models, TR 101 1112
v.3.2.0, April 1998.
[8] 3rd Generation Partnership Project, RLC Protocol Specification, 3G TS 25.322,
December 2001.
114
B. R. Vojcic is with the Department ofElectrical and Computer Engineering, The George
Washington University, Washington, DC 20052 USA {e-mail: vojcic@seas.gwu.edu}.
Abstract: This paper investigates the impact of routing on achievable SNIR and power
utilization assuming a Cellular Multihop Network (CMHN) architecture with
code division multiple access (CDMA). To support the investigation a new
centralized routing algorithm is developed to combine network layer routing
decisions with physical layer characteristics. Results are presented showing
significant advantage of the algorithm over traditional least-cost routing when
lognormal Shadowing and fading is present, demonstrating the advantage of
more sophisticated routing to maximize Signal-to-Noise-and-Interference
Ratio (SNIR) and minimize power consumption.
1. INTRODUCTION
via path diversity. However, there are unique issues associated with the
CMHN architecture. In particular, since all data is flowing to/from a
common Base Station CBS) there is the potential for considerable congestion
near the BS. Also, the required node complexity, in terms of other-user
traffic that must be handled, varies widely as a function of distance from the
BS. Finally, unlike a Single Hop (SH) network, a node must be able to both
transmit and receive on the forward link as well as the reverse link. This
paper concentrates on the impact routing has on the network interference and
power characteristics and develops a heuristic algorithm to select routes that
balance between these characteristics. It achieves this by making network
layer routing decisions based on physical layer conditions. As a result,
considerable power savings are achievable while still maximizing the SNIR.
Section II provides more detail about the CMHN architecture. Section III
analyzes the SNIR and power characteristics of a multihop network. Section
N presents a new O(K4) routing algorithm to balance between maximal
SNIR and minimum total power utilization. Section V presents simulation
results, and Section VI presents conclusions.
Figure 1b and the nodes labeled A, B, and C connected to the BS. Each of
these nodes carry traffic for themselves and 2, 4, or 5 other nodes,
respectively. Note that the cumulative sum of the traffic on the last hop
necessarily equals that of all the nodes in the network. In the example, we
have 3 packets (A~BS) + 5 (B~BS) + 6 (C~BS) = 14 packets being
transmitted to the BS on the last hop - the same as the total number of non-
BS nodes in the network. Therefore, the interference on the last hop is at
least as great as that experienced at the BS in a SH architecture.
Figure I. Example cell communications for (a) Single Hop (left) and (b) Multihop (right).
We put few restrictions on the nodes' transmit and receive abilities, except
that a node cannot simultaneously transmit and receive. We assume CDMA
as the multiple access method (although many of the results shown are
equally applicable to time or frequency division multiple access) and each
user's packet has a different spreading sequence (see [2] for code assignment
protocols). Each node is capable of simultaneously transmitting to multiple
nodes and can transmit multiple packets from different users on each link,
where the transmit power associated with each link can be separately set.
Each node can also simultaneously receive from multiple nodes. We treat
the interference caused by other users as independent additive noise ([3-5])
and assume a receiver has a required SNIR to meet performance
requirements. The route for a CMHN is a spanning tree so that on the
I The approach taken in [I] allowed intra as well as inter-cell routing. This mayor may not
be justified, depending on the type of traffic supported by the network.
118 Kevin M Pepe, Branimir R. Vojcic
forward link there are I-to-many transmissions and on the reverse link there
are many-to-I transmissions, with a total of K -1 links (K nodes includes
the BS). Importantly, since all traffic in a CMHN is routed to/from the BS,
unlike an ad hoc network, we can apply variants of more traditional wired-
network routing algorithms.
3. ACHIEVABLE SNIR
Let there be M transmitting nodes and L receiving nodes and at any given
time node i is transmitting to receiving node Ci. The routing is assumed to
Cellular Multihop Networks and the Impact of Routing 119
Note that we include the (m i -1) other packets on the same link as
interference. While it would seem pessimistic since orthogonal codes could
be used to remove the interference, in general the effects of multipath may
severely reduce orthogonality (see [14] Section 6.7.3). Furthermore, we
have not assumed any multi-user detection capability at the receiver. Thus,
(m i -1) interfering packets is a worst-case assumption.
We also note that while we derive the following equations assuming a
general set of transmitters and receivers, for the CMHN we ar~ imposing a
spanning tree route. Thus, if there are K nodes in the networlC{including the
BS) there are K-l links. Also, we treat the power setting for each link
separately. Therefore, we effectively have K-l transmitters and K-l
receIvers.
For successful communication it is assumed there exists a threshold value
that the SIR must exceed in order to support each user's required data
rate R i • For a COMA system, the SIR is usually quite small due to the
processing gain and powerful coding. To make the role of processing gain G
more explicit we can use the relationship SIR = E b / I 0 (<; W/ R t 1 , E b the
energy per bit, 10 the interference power, <; a constant of order unity which
depends on the cross-correlation properties of the spread-spectrum codes
(assumed to be one hereafter), and Wthe overall available bandwidth.
Thus for successful communication of a particular user on link ~,Ci] we
want
120 Kevin M. Pepe, Branimir R. Vojcic
(1)
or
m.r(j, c.]
where A is the MxM normalized path-gain matrixAi,j == } [. ]' Note
r I,C i
that the diagonal of A; i = mi . We see that the elements of A are
proportional to the ratio of path loss experienced by interfering transmitters
at the desired cell site, to the path loss experienced by the source transmitter
to the desired cell site. Since the transmitted power is related to the inverse
of this ratio, we see that the elements of A quantify the gain transmitter i
Cellular Multihop Networks and the Impact ofRouting 121
M M M
M'r~ IAi =tr(A)=IAii =Imi ~M~r~l,
i=1 i=l ' i=l
assuring a feasible solution. Note that the closer r gets to 1 the larger the
supportabley , assuming fixed G, or the smaller the required G (and hence
bandwidth), assuming fixed y. Thus, small r is desirable.
Finally, for CMHN we must impose the limitation that a node cannot
transmit and receive at the same time. We address this by segregating link
2 A matrix T is non-negative irreducible if all elements are non-negative and there exists a
positive integer n such that each element of Tn is strictly positive (need not be the same
n for each matrix element).
3 See Corollary 4.2 and Theorem 4.4 in Section 1.4 of [16].
122 Kevin M. Pepe, Branimir R. Vojcic
transmissions into two time slots based on whether the number of hops to the
BS is even or odd. Then A is similar (and therefore has the same
- [Ao Ae01where Ao
characteristic equation) to a matrix of the form A = 0
and Ae are normalized path-gain matrices associated with links for the odd
and even time slots, respectively. The maximal eigenvalue of A is then the
maximum of the maximal eigenvalue of Ao and A e , where Ao and Ae
are both irreducible matrices.
Note that the segregation of the channel into two time slots imposes a 3 dB
penalty to the CMHN (or any multihop versus single hop approach). That is,
for a multihop architecture, half of the time available to a node for
supporting the forward (or reverse) link is allocated to transmitting, and half
to receiving. Therefore, the power dedicated to moving information towards
its destination is reduced by half.
(13I-A)q = b
where 13 = G/y + 1, I is the MxM identity matrix, A is the same MxM path-
Cellular Multihop Networks and the Impact ofRouting 123
The analysis above indicates that for a given set of nodes and traffic
requirements and known path attenuation characteristics we can apply a
given route and determine the achievable SNIR and total network power.
Our goal in this section, and the main result of this paper, is to use the above
analysis as part of an algorithm that finds routes that balance between the
maximum achievable SNIR and the total network power. To achieve this
124 Kevin M. Pepe, Branimir R. Vojcic
4. Select Permanent Node. Select the temporary node with the smallest
label value and mark it as permanent and set p equal to this most
recent permanent node. If several nodes qualify then select the node
with the least total network power PTN •
5. Re-compute Labels. Update the labels of all temporary nodes based
on the most recent permanent node selection.
6. Check if Done. If there are no more temporary nodes then stop.
Otherwise, go to Step 2.
we must recalculate the labels and therefore incur O(k?) computations since
the matrix is now kxk. Thus, each iteration is dominated by K- k, O(e)
computations. Summing over all iterations we have
Lf=l (K -k )k 2 = O(K4 - K2 ) = O(K4) computations as the computational
complexity of the BSP algorithm.
5. SIMULATION RESULTS
distributed random variable with parameter C = 1/../2 (to ensure the average
received power is not biased), and ~ accounts for shadowing as a Gaussian
random variable with zero mean and standard deviation a .
Figure 2) shows the average SNIR and total network power characteristics
as a function of a without shadowing, and Figure 3) presents the results for
the same parameters with a = 8 dB shadowing. The SNIR is normalized to
the maximum SNIR of -1010g(K -2), which does not include the 3 dB
penalty for multihop. Note that the power level for a = 1 corresponds to the
SH power. The performance shown is typical in that for a near 1, the SNIR
is minimally impacted while still achieving substantial power savings over
the SH solution (a = 1). We see that the least-cost routing performs well
when no shadowing is present. It achieves good SNIR (within about 2 dB of
SH) while using about 10 dB less power than the single hop solution. The
BSP achieves higher SNIR for almost all values of a at the expense of about
3 dB more power than the least-cost route (but still -7 dB less power then
SH). However, when shadowing is present, as shown in Figure 3, on
average the BSP route is significantly better. For a wide range of a the BSP
route achieves greater SNIR with less power. With 8 dB shadowing and
a = 0.9 we achieve on average 9 dB better SNIR than least-cost routing and
within -0.5 dB of optimal SNIR while using 4 dB less power than the least-
cost route and 15 dB less power then SH.
Figure 4) shows the performance for 20 nodes and fixed a = 0.9 as a
function of shadowing standard deviation (in dB) without fading. We see
that the BSP algorithm's SIR performance is essentially independent of
shadowing, whereas, the greater the shadowing the better the algorithm is
128 Kevin M. Pepe, Branimir R. Vojcic
i: 8
v-
---------------------------------------------------------------
I _.... SSP r
z ·3 Route
least-cost Route
c·
.4 L--::L---::'=--_="=-_:-L-_-:"::-_-:'-:-~~2=1n::::od:::es N=o=S=ha;:::dOW\='=ng~
o ~ u ~ u ~ u U U M
alpha
Power Characteristics
.251-,---.--.,---.---y--.--;=~~;:r:;:====il
·30
I SSP Route
--- least-cost Route
21 nodes, No Shadowing
'"
E
~ ·35
~ T-----------------------------------------------------------
-40
-450L---::0~.1-~0~.2-~0~.3-~0~.4~-0~.5~~0.6~~0.7~~0.~8-~0~.9-~
alpha
SIR Characteristics
0:
Z
C/)
-4
i!
1
-6
z -8
_________________________________________ jI __ .. 21~;:st~c~~e Route
nodes. 8 dB ShadO\\ing
-10
o ~ ~ ~ U M M ~ M M
a~ha
Power Characteristics
-20 1-.-----,---,-----,------,----rr==E;;::::;;:~====il
BSP Route
-25 1 ....... least-cost Route
21 nodes, 8 dB Shadowing
-30
i -35 ----------------------------------------------------------------
~
:::~ o ~ ~ ~ U M M ~ M M
alpha
SIR Characteristics
L: -_I-----~:=----------------I------------~------.~-----._
~ --- Least-cost Route - ____ _
~ -15 20 nodes, Shadowing & No Fading
-200~----=------'-----:------:-----1:'::0------,J12
Sigma (dB)
Power Characteristics
_I
-25!r==-=;;;:;~====:::r:==~c::;==:.::.::.....-.------,--1
BSP Route
--- Least-cost Route
I
-30 20 nodes, Shadowing & No Fading
~ -35 -----------------------------------------
J -40 --------------------
-450L-----=------'----~---~----1~O---~12·
Sigma (dB)
Figure 4. Comparison ofBSP and Least-cost SNIR and Power characteristics versus
shadowing (no fading) for a fixed at 0.9.
130 Kevin M Pepe, Branimir R. Vojcic
SIR Characteristics
I:: --;----~:=--------------i----------------------
8 -15
'ij --- Least-cost Route
20 nodes, Shadowing & Fading ............. ...
z ,
·20'---------"-------"-----'----'------'-----'
o 10 12
Sigma (dB)
Power Characteristics
.251r==~~====='===:::;-,----,---r--1
1
___ ~::st~c~~eROut. I /"------------
20 nodes, Shadowing & Fading .., ....",'
S~ ~"
S -
~ -35 1......,.--:..::.--::-::.:--;..---------------------------------"-""-"-""__
-40 L __-'---__--"--__-'-__-'--__==::::::::::=:l
o 10 12
Sigma (dB)
Figure 5. Comparison ofBSP and Least-cost SNIR and Power characteristics versus
shadowing with fading for IX fixed at 0.9
6. CONCLUSIONS
This paper considered a CMHN architecture using CDMA and
investigated the impact of routing on the SNIR and power consumption
characteristics. We developed a route cost based on the SNIR and power
consumption characteristics and presented a new polynomial complexity
centralized routing algorithm to balance between SNIR and total network
power consumption. The algorithm shows that route selection has a
significant impact on the achievable SNIR and power consumption of the
network and that significant SNIR and power benefits are achievable,
especially when lognormal shadowing and fading are present. Hence,
considerable advantage can be achieved by combining network layer routing
decisions with physical layer characteristics. Furthermore, routing decisions
that do not account for the physical layer may result in significantly
degraded performance.
Cellular Multihop Networks and the Impact of Routing 131
7. REFERENCES
15) E. Seneta, Non-negative matrices and Markov chains, 2nd ed. ed. New York: Springer-
Verlag, 1981.
16) H. Minc, Nonnegative matrices. New York: Wiley, 1988.
17) N. Bambos, "Toward power-sensitive network architectures in wireless
communications: concepts, issues, and design aspects," IEEE Personal
Communications, vol. 5, pp. 50-9, 1998.
18) R. H. Jones and N. C. Steele, Mathematics in communication theory. Chichester, West
Sussex, England New York: E. Horwood; Halsted Press, 1989.
19) 1. A. Bondy and U. S. R. Murty, Graph theory with applications. New York: American
Elsevier Pub. Co., 1976.
20) W. Bunse, "A class of diagonal transformation methods for the computation of the
spectral radius of a nonnegative irreducible matrix," SIAM Journal on Numerical
Analysis, vol. 18, pp. 693-704, 1981.
21) W. H. Press, Numerical recipes in C .' the art of scientific computing, 2nd ed. ed.
Cambridge [Cambridgeshire] ; New York: Cambridge University Press, 1992.
Terminal Migration Model in Which Cell Dwell Time
is Defined by Different Probability Distributions in
Different Cells
Key words: terminal migration model, cell dwell time, mobile terminal, probability
distribution, mobile communication networks
1. INTRODUCTION
The rapid increase in the use of cellular telephones and the great
improvements of mobile phone services have created a demand for networks
that can provide various kinds of fast and stable mobile multimedia
communications services to users moving at high speeds. Because a terminal
can move from one cell to another during a call, it is necessary for the
designers and operators of such networks to consider the mobility
characteristics of terminals when they evaluate the network teletraffic. There
133
X. Lagrance and B. labbari (eds.),
Multiaccess, Mobility and Teletrajfic for Wireless Communications, Volume 6, 133-142.
© 2002 Kluwer Academic Publishers.
134 Hirotoshi Hidaka, et al.
We chose to evaluate the cell dwell time for taxis driving in the
Kurihama district of Yokosuka in Japan, because it is the largest body of
data we have (about 60 days). We analyzed the 5-km-square area the center
of which is Kurihama Station, dividing the area into cells 200 or 1,000 m
square and calculating the dwell time in each cell.
Terminal Migration Model 135
Figures l(a) and l(b) show the mean and standard deviation of cell dwell
time for each cell when the cell size is 200 m. We can see that both the mean
and standard deviation are large in the cells surrounding stations, presumably
because taxi drivers there are waiting for passengers or taking a break. On
the general roads between stations, on the other hand, the cell dwell time is
small because the taxis are only passing through the cells there. Figures 2(a)
and 2(b) show the mean and standard deviation of cell dwell time when the
cell size is 1,000 m. The difference from cell to cell obviously becomes
smaller when cells are larger. As described above, in the circumstance of
micro-cell configuration, the mean and standard deviation of cell dwell time
would differ greatly between cells.
(a) Mean
Kita-kurihama Station
(a) Mean
Up to now, the cell dwell time of a terminal has been modeled by one
probability distribution having the same mean and standard deviation in all
the cells. In particular, an exponential distribution, which is easy to
understand the phenomenon and calculate, has been used frequently. It was
shown in Section 2, however, the mean and standard deviation of the cell
dwell time differs greatly between cells in the micro-cell configuration. This
means that the probability distribution followed by the cell dwell time also
Terminal Migration Model 137
Table 1. Mean and standard deviation of cell dwell time (sec.) for three typical cells
Table 2. Parameters of distributions approximating the cell dwell time for three typical cells
other hand, the mean is of course nearly equal to the standard deviation
because the cell dwell time was determined by one exponential distribution
for which the mean was the same in all the cells.
We also ran simulations in which the cumulative probability of cell dwell
time was approximated by exponential and log-normal distributions
estimated using the least-squares method. The results are listed in Table 5.
The results listed for the Type II simulations clearly followed a log-normal
distribution more closely than they did an exponential distribution. In the
Type III simulations, although the regression coefficient for the log-normal
distribution is large, the mean approximated by the exponential distribution
is so close to that of original data that the data follows the exponential
distribution very well. With regard to the Type I simulation, it is worthy of
attention that the results follow the log-normal distribution well though the
dwell time in each cell was given by the exponential distribution. This seems
to result from the effect of long-tailed characteristic because, in the cell A,
random exponential numbers, of which the mean is great, may be generated.
Straight 0.3
Right 0.3
Left 0.3
Reverse 0.1
Table 4. Mean and standard deviation of the cell dwell time ofa terminal in the simulation
1.0 I-------:::;~~---I
0.8
0.6
lJ..
C>
o
0.4
-0- Type I
- . - Type II
-x- Type III
10 100 1.000
Celldwelltin e [5]
Figure. 4. Cumulative distribution ofthe cell dwell time of a terminal in the simulation.
Table 5. Parameters of distributions approximating the cell dwell time for three types.
In Ref. 8 we showed that the actual cell dwell time measured for a taxi
followed a log-normal distribution rather than exponential distribution,
seemingly because the cell dwell time of taxis is large around a station and is
small in other places. In other words, the cell dwell time differs between
cells. Although the results reported in this paper may be applicable only for
taxis, cell dwell time will generally vary from cell to cell. It is therefore
necessary to analyze the mobility characteristics of other platforms for every
cell in the same way that it is necessary to analyze them the mobility
characteristics of taxis for every cell.
Terminal Migration Model 141
5. CONCLUSIONS
In the conventional model of terminal migration the cell dwell time has
been expressed by one probability distribution which is used commonly in
all the cells, assuming that a terminal moves at random independent of
locality. In actual communication systems, however, the dwell time in each
cell is expected to differ greatly. This paper analyzed the cell dwell time of
taxis we have measured so far and it showed that the mean and standard
deviation of cell dwell time are large in certain places, such as those within
or close to train stations, and are small in cells along general roads. It also
showed that the dwell time in each cell followed a log-normal distribution
rather than a conventionally used exponential distribution.
We also considered a migration model in which the cell dwell time is
defined by the different probability distribution from each cell, and we used
computer simulations to evaluate the cell dwell time of a terminal. Even if
the dwell time in all cells was given by an exponential distribution, the
results of those simulations showed that the cell dwell time followed a long-
tailed log-normal distribution. Although these results may be applicable only
to taxis, the methodology described in this paper-taking account of cell
dwell time following different probability distributions in different cells-
offers a realistic and effective way to evaluate teletraffic characteristics in
future mobile multimedia communication systems.
REFERENCES
[1] D. Hong and S. S. Rappaport, "Traffic model and performance analysis for cellular mobile
radio telephone systems with prioritized and non-prioritized handoffprocedures," IEEE
Trans. Veh. Tech., vol. VT-35, no. 3, pp. 77-92, Aug. 1986.
[2] P. Orlik and S. S. Rappaport, "Traffic performance and mobility modeling of cellular
communications with mixed platforms and highly variable mobilities," Proc. IEEE, vol. 86,
no. 7, pp. 1464-1479, July 1998.
[3] Y. Fang and I.Chlamtac, "Teletraffic analysis and mobility modeling of PCS networks,"
IEEE Trans. Commun., vol. 47, no. 7, pp. 1062-1071, July 1999.
[4] M. M. Zonoozi and P. Dassanayake, "User mobility modeling and characterization of
mobility patterns," IEEE 1. Select. Areas Commun., vol. IS, no. 7, Sept. 1997.
[5] F. Khan and D. Zeghlache, "Effect of cell residence time distribution on the performance
of cellular mobile networks," Proc. IEEE VTC'97, pp. 949-953, May 1997.
[6] K. K. Leung, W. A. Massey, and W. Whitt, "Traffic models for wireless communication
networks," IEEE J. Select. Areas Commun., vol. 12, no. 8, pp. 1353-1364,1994.
[7] Bar-Noy and I. Kessler, "Mobile users: to update or not to update?," Proc. INFOCOM'94,
pp. 570-576, 1994.
142 Hirotoshi Hidaka, et ai.
1. INTRODUCTION
2. PROBLEMS
With this approach, a group of MTs that moves simultaneously into a new
location registration area tries to register their location simultaneously as
shown as Fig. 1. This is a serious problem because it is impractical to
provide sufficient resources to handle the flood of signals thus generated.
Since we must anticipate that many MTs will share the same moving
platform, such as a train, they will have the same movement characteristics,
and a more effective location registration approach appears possible.
Therefore, we propose Concatenated Location Management (CLM), which
enables location information to be updated without individual location
registration.
The logical solution is to equip the train with an Intermediate Radio
Station (IRS) and update the location information of all MTs with one action.
This relationship between the MT and the IRS is called "Concatenation" in
this paper. There are two ways of implementing this concatenation: IRS
(method 1) and the network (method 2).
In method 1, the IRS updates the location information of all MTs that are
traveling with the IRS. In method 2, the MTs update their location
information to indicate that they are traveling with the IRS when they join
the IRS. The network is responsible for updating the location information of
the MTs moving with the IRS.
Fig. 2 shows the basic procedures of method 1. The details of method 1
are as follows.
Fig. 3 shows the procedures for method 2. The details of the method are as
follows.
MT#2;
location regim-alion ar~a A
·>concatenal;ng to IRS #1 (3)
4. PAGING CONTROL
In the conventional method, the network pages an MT from all BSs in the
registered location registration area of the MT. The network identifies the
BS where the MT visits when the MT responds to the paging signal. In this
section, we propose paging control for CLM.
The conventional method has the network page the MT directly, and this
idea can also be adopted for CLM (alternative 1: AI). An alternative is for
the network to notify the IRS of the paging request, and have the IRS page
the MT. There are two responses possible; one is that the MT responds to the
148 Koji SASADA, Satoshi HIYAMA, and Masami YABUSAKI
network directly (alternative 2: A2), and the other is that the MT responds to
the network via the IRS (alternative 3: A3). Fig. 4 shows each alternative.
There are two ways to manage location information: Bl and B2. In Bl,
the network manages the conventional location registration area of the MT.
In B2, the network identifies the location registration area through the
concatenation relationship.
B 1 requires only one database that holds the location information of the
IRS and all MTs as shown in Fig. 7. This enables the network to identify the
MT's location registration area from the table and page the MT directly. The
network does not manage the concatenation between the MT and IRS so
there is no need to notify the location register about changes in the
concatenation relationships.
B2 requires two databases as shown in Fig. 7. This figure shows that MT
#1 is concatenated to IRS #1, and that IRS #1 is in location registration area
A. Clearly MT #1 is visiting location registration area A. The network uses
the concatenation relationship when paging the MT. Note that the location
register must be informed of any change in the concatenation relationships.
location infonnation
location table for Mrs and IRSs
concatenating to IR...o;; #1
10 location infonnation
location registration area B i
MT #1 location registration area A
i················································..
.,i
MT #2 location registration area B
IRS #1 location registration area A location infonnation
locationrc&istcr
location registration area A
6. RESULT
result is that eLM method 2 could reduce the number of signals by 91.2%
compared to the conventional location registration scheme (Fig. 10). This is
because the express runs long distances with few stops.
1.400
1.200
1.000
800
600
400
200
o
Shin- Okayama Hiroshima KoItUlO Hakala
ToIt)'O Yoltohama Nagoya Kyoco
0.0 28.8 366.0 513.6 552.6 732.9 894.2 1107.7 1174.9
Distan<. from slart (Icrn]
20.000
18.000
16.000
-il14.ooo
Iii.
;: 12.000
r::
o
6,000
4,000
2,000
o
s.....
KjOIO ow. Obyamo W..... ima K<>I<u.. llabla
8. CONCLUSIONS
REFERENCES
[1] H. Yumiba et ai, "IP-Based IMT Network Platform," IEEE Personal Communications,
pp.18-23, Oct. 2001.
[2] "Location registration procedures (GSM 03.12 version 7.0.0 Release 1998)," ETSI, Aug.
1999.
[3] "Location Registration Control," Personal digital cellular telecommunication system
ARIB standard RCR STD-27H, pp.889-895, Feb. 1999.
[4] "General Packet Radio Service (GPRS) Service description (3GPP TS 23.060 version
3.10.0 Release 1999)," ETSI, Jan. 2001.
HANDOFF SCHEME IMPROVEMENT IN WIRELESS
NETWORKS
Abstract
Chaining followed by make-break algorithm reduces signaling traffic in
wireless ATM network and improves the efficiency of the virtual channel. We
propose a new handoff call management scheme which complies with existing
ATM signaling standard and its implementation leaves commercially ATM com-
ponents unaffected. This method considers traffic condition when chaining and it
is easy to implement in the chaining followed by the make-break scheme.
Keywords:
wireless ATM networks, routing, handoff algorithms.
1. INTRODUCTION
Wireless data services use small-coverage high-bandwidth data networks
such as IEEE 802.11 whenever they are available, and switch to an overlay ser-
vice such as the General Packet Radio Service (GPRS) network with low
bandwidth when the coverage of a wireless local area network (WLAN) is not
available [3].
From the service point of view, ATM (Asynchronous Transfer Mode) com-
bines both the data and multimedia information into the wired networks while
scaling well from backbones to the customer premises networks. In wireless
ATM (W ATM) networks, end user devices are connected to switches via wired or
wireless channels. The switch is responsible for establishing connections with the
fixed infrastructure network component, either through wired or wireless channel.
A mobile end user establishes a virtual circuit (YC) to communicate with another
end user (either mobile or ATM end user) [I]. When the mobile end user moves
from one AP (access point) to another AP, a handoff is required. To minimize the
interruption of cell transport, an efficient switching of the active YCs from the old
data path to the new data path is needed. Also the switching should be fast
enough to make the new YCs available to the mobile users.
When a handoff occurs, the current QoS (Quality of Service) may not be
supported by the new data path. In this case, a negotiation is required to set up a
new QoS. Since a mobile user may be in the access range of several APs, it will
155
X. Lagrance and B. labbari (eds.),
Multiaccess, Mobility and Teletrafficfor Wireless Communications, Volume 6,155-166.
© 2002 Kluwer Academic Publishers.
156
(i) The connection server informs the new base station of the completion of the
route change, which then starts using the new route.
(j) The connection server exchanges messages with the ATM switch, removing
the old routing entry. The connection server also requests the old and new
base stations and switches in the old route to release the old resources.
the entire route is improved. As the result, the chance of going through steps (h)
and (i) is reduced so is the signaling traffic involved in (h) and (i).
Signaling traffic depends on the network configuration and protocols
involved. When a mobile user roams within the A TM switch area, the signaling
traffic is low, and it performs according to the Chain Routing algorithm scheme.
When a rerouting process is required, the signaling messages are a few bytes long,
because only one ATM switch is involved. The longest message is the handoff
request message from the mobile user. This message is 44 bytes long and
includes the mobile identity, old base station channel server identifier, and the 3-
tuple (VPI, vcr, and port) of the translation table entry at the same ATM switch.
The route update message to the connection server contains the identity of the
mobile endpoint and two base stations involved in the Chaining scheme.
When the mobile user roams outside the original ATM switch and a reroute
is requested because of the overload of links or QoS problem, the new base sta-
tion needs to identify the best route to the crossover switch, allocate resources
along the route, and then exchange messages with the crossover switch, which
executes break-make or make-break operations. In this case, Chain Routing algo-
rithm performs better than the Chaining scheme.
1. In certain cases, because of the Chain Routing algorithm, the links connect-
ing base stations, and the links connecting A TM switch and base station are
used more efficiently, thus this re-route does not occur as often as the
Chaining scheme.
2. In certain cases, when the mobile user roams to the other A TM switch area,
the chain part inside the original A TM switch area will be rerouted accord-
ing to the Chain Routing algorithm, thus this portion of the routing path
will not likely have overload problem and the QoS problems as the Chain-
ing scheme does, and the overall routing path is not likely to be rerouted as
it is in the Chaining scheme.
The difference can be demonstrated by the different call drop rates in cer-
tain network configurations.
sent, whereas only one operation (make) is needed in the make-break scheme.
The Chaining scheme is fast because it preassigns VCs between neighboring base
stations and, thus, translation entries at the crossover switch need not be to
changed. If VC's were not preassigned, the handoff latency in the Chaining
scheme would be comparable to that of the make-break scheme.
Chaining with break-make and Chaining with make-break perform their
rerouting operations after handoff and, thus, those operations do not affect the
handoff latency of these schemes. Note that the handoff latency measurements
depend on the number of connections of the mobile endpoint that must be
rerouted. This is because each connection corresponds to a translation table entry
in the switch. Therefore, rerouting mUltiple connections implies that multiple
translation table entries have to be modified, resulting in higher latencies.
Regarding the impact of connection rerouting involving multiple ATM switches
on handoff latency, the handoff latency in Chaining with break-make and Chain-
ing with make-break will not be affected since latency is determined only by the
chaining of the neighboring base stations. On the other hand, in break-make
scheme and make-break scheme, handoff latency is directly proportional to the
number of ATM switches that need to be updated along the new route. Thus, the
separation of connection rerouting from the real-time phase in schemes Chaining
with break-make, and Chaining with make-break results in a low hand off latency
regardless of the number of switches involved in the rerouting operations. In the
Chain Routing algorithm scheme, the Chain Routing algorithm only applies to the
chain part of the route path, translation entries at the crossover switch need not to
be changed. The time cost of chain routing attributed to handoff latency will be
comparable to the Chaining algorithm.
In certain cases, some calls are blocked because of the overload of chain
part of the route path and those links between ATM switch and base stations. In
cases that those calls have to been rerouted, the handoff latency is comparable to
make-break scheme or break-make scheme. In the Hop-limited handoff and
Chain Routing algorithm schemes, the routes have to be rerouted at certain cir-
cumstances.
In the Hop-limited handoff scheme, when a mobile has successfully made
r-l handoffs, and its rth handoff request is also successful, its traffic path would
be rerouted from the new base station to the ATM switch to which it belongs.
Regardless whether the mobile user roams out of the current ATM switch area to
the new ATM switch or not, the connection server performs make-break or
break-make and necessary QoS computations on the new route. If the mobile user
roams out of the current ATM switch area, the handoff latency is directly propor-
tional to the number of ATM switches that need to be updated along the new
route.
The handoff latency is similar to make-break or break-make scheme. For
the Chain Routing algorithm scheme, the route will be rerouted when the occu-
pancy of the VCs between current base station and its neighboring base station or
of the VCs between ATM switch and base station reach a certain value. Because
two cases are considered, one is that the user roams inside an ATM switch area,
and the other is that the user roams outside the current ATM switch area, and the
handoff latency is different from the Hop-limited handoff scheme. Regardless
164
whether the user roams inside an ATM switch area or the user roams outside the
current ATM switch area, only the chain part of the route path inside the original
ATM switch area will be rerouted. The handoff latency is similar to the Chaining
scheme. Because handoff latency of the Chain Routing algorithm consists of
rerouting cost and chaining cost and handoff latency of the Chaining scheme con-
sists of chaining cost only, the latency of Chain Routing algorithm is higher than
of the Chaining scheme. Depending on the r value of the Hop-limited handoff
scheme, the latency of Chain Routing algorithm is higher than of the Hop-limited
handoff scheme with a large r value but shorter than of the Hop-limited handoff
scheme with a small r value. The routes that are blocked have to be rerouted as
those in break-make and make-break schemes.
Regarding the impact of connection rerouting involving multiple ATM
switches on handoff latency, the handoff latency in Chaining with make-break
will not be affected since latency is determined only by the chaining of the neigh-
boring base stations. Thus, the separation of connection rerouting from the real-
time phase in Chaining with make-break results in a low handoff latency regard-
less of the number of switches involved in the rerouting operation. While connec-
tion rerouting due to handoffs is similar to rerouting due to the failure of network
components, thcre are two important differences. First, handoffs are much more
frequent than network faults. With frequent reroutes, the disruption caused to
ongoing connections has to be minimized. On the other hand, in many cases
applications will be willing to tolerate some disruption due to rare network fault
rerouting scenarios. Second, handoffs result in connection reroutes that are lim-
ited to a small geographic locality (e.g., neighboring base stations). On the other
hand, reroutes due to failures may involve reestablishing the entire connection.
In ATM networks, all data is transmitted in small, fixed-size packets. Due
to the high-speed transfer rate (in the range of hundreds to thousands of Mb/s) and
short cell length (53 bytes), the ratio of propagation delay to cell transmission
time and the ratio of processing time to cell transmission time of ATM networks
will increase significantly more than that in the existing networks. This leads to a
shift in the network's performance bottleneck from channel transmission speed (in
most existing networks) to the propagation delay of the channel and the process-
ing speed at the network switching nodes. This chain routing method will
decrease the work load in the network switching nodes.
1. There is no need to identify the crossover switch when rerouting, because
chain routing method works only with one ATM switch.
2. There is no need to calculate the best route through the connection server
because it is done locally to reduce signaling traffic.
3. It is easy to implement. Only one parameter ORP is added to the new
scheme, and the calculation is very simple. It complies with existing ATM
signaling standard, and its implementation leaves commercially available
ATM components unaffected.
4. It is in real time.
5. It can significantly reduce signaling traffic.
6. In the new handoff scheme, the concern of traffic jam is included. This
scheme can handle different kinds of situations efficiently. By doing that,
165
the entire PVC in this ATM switch will have the highest utility efficient, so
that system adopting this scheme can handle much more handoffs.
4. CONCLUSION
The signaling traffic is significantly reduced by using the Chain Routing
algorithm scheme or the Hop-limited handoff scheme. The Hop-limited handoff
scheme or the Chain Routing algorithm scheme should be added to the Chaining
and make-break schemes depending on different situations. These schemes
significantly reduce the signaling traffic in the network, which causes lower
number of call drops, and there is no need to check QoS often, and to re-route
often, in the chaining part of the route path. Without the two methods, more
checking or rerouting needs to be done, with more signaling traffic in the W ATM
network. These methods significantly reduce the number of call drops in the
chaining parts. Chain Routing algorithm scheme is more suitable when the PYCs
between the base stations and the ATM switches, and the PYCs between the base
stations are very limited or there is a rush of handoff requests in the W ATM net-
work.
The Hop-limited handoff scheme is more suitable when the PYCs between
the base stations and the ATM switches, and the PYCs between the base stations,
are not limited, and the traffic in the W ATM network is not heavy. Depending on
different situation, different r values should be selected.
Configuration of the network does not make much difference, but the
number of cells inside one ATM area. When selecting a scheme, one factor must
be considered that overhead of Chain Routing algorithm scheme is higher than the
Hop-limited handoff scheme because it needs to check the ORA when it decides
if rerouting is needed.
Chaining with make-break is ideally suited for performing connection
rerouting, Chain Routing algorithm scheme or the Hop-limited handoff scheme
make the best use of it. The limitation of Chain Routing algorithm scheme is that
it has no effect on the mobile users coming from the other ATM switches.
5. REFERENCES
1. Chan, K.S. Hop-limited handoff scheme for ATM-based broadband cellular
networks. Electronics Letters 1998; 34:26-27
2. Ghai, R., Singh, S. An architecture and communication protocol for picocel-
lular networks. IEEE Personal Communications 1994; 1:36-46
3. Pahlavan, K. Handoff in hybrid mobile data networks. IEEE Personal Com-
munications 2000; 7:34-47
4. Ramjee, R. Performance evaluation of connection rerouting schemes for
ATM-based wireless networks. IEEE/ACM Transactions on Networking
1998; 6:249-61
Hierarchical Mobility Controlled by the Network
Abstract: The majority of IP based existing mobility solutions have significant handover
execution time and movement detection delay. In this article, we present
NCHMIPv6 (Network Controlled Hierarchical Mobile Internet Protocol
version 6), a management protocol for hierarchical mobility controlled by the
network.
1. INTRODUCTION
These last years, Internet became very popular and made IP protocol
essential for the development of telecommunications networks. At the same
time, the miniaturization of data-processing equipment like portable
computers and the development of wireless networks and services increased
the need for mobility.
An IP mobile node, connected to the Internet, is localized compared to its
attachment point with its IP address; if it moves to a new attachment point, it
must imperatively:
• change its IP address: it is for example the solution adopted by the
dynamic address allocation scheme allowing users nomadism between
several sites; this solution however requires to stop all IP transfers in
progress; it thus allows the nomadism but does not provide servIce
continuity in the course of mobility
167
X. Lagrance and B. Jabbari (eds.),
Multiaccess, Mobility and Teletrafficfor Wireless Communications, Volume 6,167-182.
© 2002 Kluwer Academic Publishers.
168 Y. Khouaja, K. Guillouard, JM. Bonnin, P. Bertin
• or inform all the routers of the Internet that its IP address locates a new
attachment point, it is an unrealistic solution.
Mobile IP [Perk97] is the standard protocol for the support of macro-
mobility i.e. mobility between networks. It allows a transparent routing of
IP packets destined for the mobile nodes in Internet and ensures
consequently service continuity for communications in progress. However,
if Mobile IP is used to manage micro-mobility, i.e. mobility localized in the
same network, it results in introducing delays in the diffusion of the new
localization and generates significant control traffic in the Internet core.
These last years, several micro-mobility protocols were developed [EMSOO]
to solve these problems. Nevertheless, the handover execution time
managed by these protocols remains always considerable and the movement
detection is delayed. We then propose in this paper NCHMIPv6, a new IP
micro-mobility protocol extended from HMIPv6 [COO], which includes the
decision-making phase of the handover execution in the mobility
management protocol and uses a handover scheme controlled by the
network.
This article consists of three principal sections. In the first section, we
briefly describe the protocol Mobile IPv4 and we develop a short
comparison with Mobile IPv6. The second section presents a classification
in two families of micro-mobility protocols: Proxy Agents Architecture
protocols and Localized Enhanced-Routing protocols. Then, the limits of
these protocols are detailed. In the third section, we present the architecture,
procedures and evaluation ofNCHMIPv6.
2. IP MACRO-MOBILITY
Mobile IPv4 protocol defines three functional entities: the mobile node,
the home agent and the foreign agent. The mobile node is configured with a
Hierarchical Mobility Controlled by the Network 169
establish a topologically correct reverse tunnel from the care-of address, i.e.
either the mobile node or the foreign agent depending on the temporary
address of the mobile node, to the home agent. Sent packets are then
decapsulated by the home agent and delivered to correspondent nodes with
the home address as IP source address.
As the home agent intercepts all packets addressed to the mobile node
and tunnels them to the visited network, a "triangle routing" effect is
produced: all packets must first pass through the home agent even if the
current access router is in the same network as the correspondent node. An
extension of Mobile IP, known as Route Optimisation [PJOI], has been
proposed to overcome this problem: it allows data packets to be routed
directly from the correspondent node to the mobile node using a binding
cache in the correspondent node that keeps track of the current temporary
address. These binding caches are created and updated by Binding Update
messages sent by the home agent or the mobile node in response to mobile
node warnings or correspondent node requests.
3. IP MICRO-MOBILITY
These protocols define the network architecture in two levels: the highest
level is standard Internet network implementing Mobile IP and the lowest
level is composed of domains able to manage micro-mobility. The router
connecting the two levels masks local movements compared to the rest of
Internet network. Inside a domain, a specific routing protocol is introduced.
The packets emitted by the mobile node update the routing entries in the
intermediate nodes. Then a routing table entry maps the mobile node address
with the address of the neighbouring node having transmitted the last packet
responsible for the entry update. The chain of these correspondences
represents the path traversed by the packets destined for the mobile node.
The most known protocols are Cellular IP [CGKTWVOO] and HAWAll
[RLTVOO].
172 Y. Khouaja, K. Guillouard, J.M Bonnin, P. Bertin
3.3 Limits
4. NCHMIPV6
4.1 Architecture
MM Mobility Manager
MN Mobile Node
AP Access Point
AR Access Router
4.2 Initialisation
\
l'
/
I /
/
/
3
. .. - . - . Signalling
1. @L -)@home. ~
MN
C Corrcspon dant ode
O. The mobile node accedes to
NCHMIPv6 domain
The virtual eOA address plays the role of Mobile IPv6 primary eOA
address. The mobile node informs its home agent and its correspondent
nodes, eN, of this eOA address by using Mobile IPv6 Binding Updates
(stage 1). Moreover it must be informed of the correspondence between the
virtual address and the local address, so that the mobility manager is able to
deliver these packets. For that, as soon as it acquires these two eOA
addresses, the mobile node informs the mobility manager of the
correspondence between the virtual address and the local address (@L, @V)
by transmitting an HMIPv6 registration request (stage 1'). Thus, all the
packets addressed to the mobile node are routed towards its virtual address
@V at the level of the mobility manager. Then, the mobility manager
intercepts these packets and delivers them to the mobile node local address
Hierarchical Mobility Controlled by the Network 175
@L. The mobility manager, after having consulted its database (stage 2),
acknowledges the registration request while returning to the mobile node a
Mobile IPv6 Binding Acknowledgement which must include, moreover, the
list of the neighbouring access points of the mobile node current access point
in a new NCHMIPv6 option (stage 3).
mapping the new local mobile node IP address to the hardware address
diffused in the neighbour advertisement. From this moment, the
communications towards the mobile node can start again since the packets
are already redirected at the level of the mobility manager. With the
reception of the router advertisement from its new access router, the mobile
node completes and updates the parameters of its new local COA address
(lifetime, configuration flags ... ).
Signalling
5
..
.... :::::....,..~..
~
'. ~
6. @L2 -+@home . ••••••
'. '. " ':::.2 MN
.... O. @L1 -+@ home.
6 S. @L l -+@bo me.
...." ,,: I
@L2 @home .
....... ........ 1
Stages of bandover management:
The mobile node must also check the nature of the movement. If it is a
movement inside the same domain, it emits an HMIPv6 registration request
towards its mobility manager in order to confirm the new binding (@Vl,
Hierarchical Mobility Controlled by the Network 177
@L2) and to eliminate the old one (@Vl, @L1) (stage 8). If it's a
movement between domains, the mobile node acquires, in addition to its
local address @L2, a new virtual address @V2 and announces it to its new
mobility manager and to its corresponding nodes by emitting respectively a
new HMIPv6 registration request and Mobile IPv6 Binding Updates. The
new mobility manager acknowledges the request by a Mobile IPv6 Binding
Acknowledgement containing the list of the neighbouring access points of
the new access point in NCHMIPv6 option.
During all the handover management phase, the mobile node continues to
monitor the radio link quality with its current access point. If quality is
degraded under a critical threshold S2, the mobile node keeps the possibility
of executing a handover towards the access point having the best radio
quality, without awaiting the handover response from the mobility manager.
This ensures a minimal quality for established communication.
Signalling
/
/ 5
2. @Ll
3. @L1
~@home.
~@home,
2 ,.... ,.. ,"""",[ ~N I o.
I
@L1 ~@home.
@L2~@home. 1
5. @L2 ~@home.
1. Quality < S 4. Neighbour Advertisement
+ Router Sollicitation
Stages of handover
management: 2. MN moves 5. Registration Request
3,5
3 ..... .-.
..--.-
<II
2,5
~... 2
<II
> 1,5
0
'C
c;;
",,- ;
(II
:c:
1 ,
• • • • • • • • •
0,5
0
C)~ C)~
C), C),
,,~ C)'r "v~
C),
c)'? C"J~) , "
,,~ "bi,~
v
Transmission time
5. CONCLUSION
REFERENCES
[COO] Castelluccia, "An Hierarchical Mobile IPv6 Proposal", INRIA, November 1998.
[BOO] Bellier, http://www.inrialpes.fr/planete/people/bellier/hmip.html.
[CGKTWVOO] Campbell, A., Gomez, J., Kim, S., Turanyi, Z., Wan, C-Y., Valko, A.,
"Design, Implementation and Evaluation of Cellular IP", Communication IEEE, July 2000.
[DH98] Deering, S., Hinden, R., "Internet Protocol, Version 6 (IPv6) Specification", RFC
2460 IETF, December 1998.[Drom97] Droms, R. ,"Dynamic Host Configuration Protocol",
RFC 2131 IETF, March 1997.
[Dupont99] Dupont, F., ftp:/Iftp.inria.fr/networklipv6/.
[EMSOO] Eardley, P., Mihailovic, A., Suihko, T., "A Framework for The Evaluation of IP
Mobility Protocols", In Proceedings ofPIMRC 2000, London, UK, September 2000.
[FS98] Ferguson, P., Senie, D. ,"Network Ingress Filtering: Defeating Denial of Service
Attacks which employ IP Source Address Spoofing", RFC 2267 IETF, January 1998.
[GJPOI] Gustafsson, E., Jonsson, A., Perkins, C., "Mobile IP Regional Registration", Draft
IETF, September 200 I.
[IEEE97] IEEE Std 802.11-1997, "Wireless LAN Medium Access control (MAC) and
Physical Layer (PH) Specification".
[JPOI] Johnson, D., Perkins, C.,"Mobility Support in IPv6", Draft IETF, July 2001.
[MontOI] Montenegro, G. ,"Reverse Tunneling for Mobile IP, revised", RFC 3024 IETF,
January 2001.
[NNS98] Narten, T., Nordmark, E. Simpson, W., "Neighbor Discovery for IP Version 6
(IPv6)", RFC 2461 IETF, December 1998.
[Perk97] Perkins, C., "Mobile IP", IEEE Communications Magazine, May 1997.
[PerkO I] Perkins, C., "IP Mobility Support for IPv4, revised", RFC 2002, September 200 I.
[PJOI] Perkins, C., Johnson, D., "Route Optimization in Mobile IP", Draft IETF, September
2001.
[RLTVOO] Ramjee, R., La Porta, T., Thuei, S., Varadhan, K., "IP micro-mobility support
using HAWAII", Draft IETF, July 2000.
[SCEBOI] Soliman, H., Castelluccia, C., El-Malki, K., Bellier, L.,"Hierarchical MIPv6
mobility management", Draft IETF, July 200 I.
Approximate and exact ML detectors
for CDMA and MIMO systems:
a tree detection approach
abstract This paper deals with Direct-Sequence Code Division Multiple Access
(DS-CDMA) transmissions over mobile radio channels. Different detection tech-
niques have been proposed in the past years, among which approximate ML detec-
tors. In this paper we propose an exact ML detector with a low complexity. We show
that a QR factorization of the matrix of users' signatures is appropriate: the upper
triangular form of the R matrix makes it possible to state the ML detection in terms
of a shortest path detection in a tree diagram. Different algorithms based on the tree
diagram are compared: the stack algorithm (exact ML), and the feedback decoding
algorithm (approximate ML). The numerical complexity of the proposed techniques
is studied in detail; in particular, the low complexity of the stack algorithm at high
SNR is pointed out. Simulations show that the performance of the stack algorithm in
terms of Bit Error Rate (BER) are very close to the single user bound. Furthermore,
we point out the fact that the same kind of approach can be used to perform ML
detection in Multiple Input Multiple Output (MIMO) systems.
1 Introd uction
Until the mid 80's, the techniques that were used to detect symbols trans-
mitted by a multiuser DS-CDMA system were based on single user detection
strategies. Unfortunately, this kind of approach yields very poor performance
due to the lack of orthonormality of the signals arising from distinct users.
Interference among users is known as Multiple Access Interference (MAl) and
it is all the more sensitive as the interfering users have high power, yielding
what is called the near-far-effect. In order to suppress MAl, multiuser detec-
tion strategies have been developped since 1986 [11].
In particular, MAl can be suppressed using the popular decorrelator or
linear Minimum Mean Square Error (MMSE) techniques. They are not so
efficient in terms of Binary Error Rate (BER) performance as the Maximum
Likelihood (ML) detector, but they require smaller computational effort [13].
In fact, the ML detector consists in selecting the most probable combination of
users bits among 2K , where J{ is the number of users. If an exhaustive search
is performed, ML detection is impracticable in many real systems [12]. It is
183
X. Lagrance and B. labbari (eds.),
Multiaccess, Mobility and Teletrafficfor Wireless Communications, Volume 6,183-194.
© 2002 Kluwer Academic Publishers.
184
2 Problem statment
For the sake of conciseness we will only consider real valued signals, extension
to complex signals being straightforward. We consider in this section a syn-
chronous DS-CDMA communication system where K users transmit data sym-
bols (bkh=l,K with bk = ±l. (Sk)k=l,K represent the corresponding spreading
sequences with length N (N 2: K). Then noting b = [b 1 , ... ,b K F the vector
of users bits, and S = [S1, ... ,SK] the matrix of signatures, the outputs of the
BPSK or QPSK demodulator sampled at the chip period is the vector
r = Sb+n, (1 )
(2)
In our approach the matched filtering is not necessary and we consider expres-
sion (1).
Similarly, in MIMO communications, the data are transmitted from an ar-
ray of K sensors to an array of N sensors. Denoting by r the vector observed
on the receiver array yields data model (1), where b represents the transmit-
ted data vector and S is the channel mixing matrix. This model is valid when
instantaneous, that is to say non convolutive, transmission channels are con-
sidered. Working from model (1), it is clear that the detection of b can be
adressed in the same way for CDMA and MIMO systems.
where II x 112= x T x.
Consider now the QR factorization of S,
S=QR, (4)
(5)
The second term on the right-hand side of this equality does not depend on
b and can therefore be omitted for the optimization. Problem (3) is then
equivalent to:
{ minb II Rb - z W (6)
bk = ±1, k = 1, ... ,K,
(7)
i=K,K -1, ··,1
(Rb)i is the component number i of the vector Rb; that is to say (Rb)i =
"L,j=i,K Rijbj , and (Rb)i only depends on bits bj , for j = i, i + 1, ... , I<.
Then, the vector b can be identified to one path in a tree diagram, starting
from the root with user K, then coming users number K - l,I{ - 2, ... ,1. The
criterion II Rb - z Win problem (6) can be decomposed into a sum of K
187
Userbit= "+1"1
Userbit= "-1"1
branch metrics of this diagram: the metric of the branch bk - 1 extending the
path bK , bK- 1, ... ,bk is I(Rbh-l - zk_11 2 = I~j=k-l,K Rk-l,jbj - zk_11 2 .
Figure 1 corresponds to the case [{ = 3 users. The bold path stands
for the vector of users bits b = (b 1, b2 , b3 f = (+1, +1, _1)T. Finally, the
criterion II Rb - z Wis the cumulated branch metrics along the path b =
(b K , bK -1, ... ,b 1 ) in the tree. The solution of the problem (6) is therefore the
path b with the shortest cumulated metric.
step 1
repeated until a complete path with K branchs is found. The stack algorithm
with L:::: 3 is illustrated on Figure 2.
5 Computational issue
The computational costs involved by the detection techniques that we have
just presented consist in two contributions: (i) a preprocessing that requires
the calculation of the QR decomposition of the matrix S (ii) the search of a
shortest path in a tree diagram. We are going to see that in many situations
189
5.1 Preprocessing
The computational effort required to change problem (3) into problem (6)
mainly amounts to the QR decomposition of the matrix S. It involves O(N [{2)
operations, and it can be implemented in several ways [3].
At this point, we can note that this is not more than performing the
decorrelator detection. Indeed, a direct implementation of the decorrelator
detector involves O(N [{2) operations for calculating the correlation matrix
STS of users signatures and O([{3) operations for its inversion. In fact, the
decorrelator detector is generally implemented iteratively by means of an iter-
ative Successive Interference Cancellation (SIC) detector [6, 10]. This involves
O(N [{) operations per iteration. But the number of iterations that must be
used to ensure convergence of the SIC detector to the decorrelator increases
significantly with the number of users. Note that the linear MMSE detector
can be implemented with similar computational complexity. This shows that
performing the QR decomposition of the matrix S is not significantly more
complex than considering the decorrelator or the MMSE detector.
feedback decoding algorithm requires less computation than the stack, but it
performs only approximate ML detection.
6 Simulation Results
We consider the case of K = 9 users with signatures of length N = 20 chips.
The signatures are Gold sequences of length 63, truncated at N = 20. Then,
the matrix S is of the form
S = (l/VNJx T
+-++-+--+---+--++--+
+-++-++---+++-+----+
--+--+++------+--+-+
+--+--+++--++++-----
+++++-+-+-+--++-+-++ (8)
--+-+---++-+-+++++--
+---++----++-+-+--++
++---+-+++++----++-+
+--+-+++++--++------
-+-+-++--++++-++---+
where sign + stands for +1 and sign - stands for -1. Typically, this kind
of matrix can be met in the case of a DS-CDMA communication. Note that
the absolute value of the correlation among the columns ranges from 0 to 0.5
and that the mean and the standard deviation of the off-diagonal terms of the
covariance matrix STS are given by 0.044 and 0.224.
Figure 4 shows a comparison of the numerical complexity of the stack
algorithm and the feedback decoding algorithm. For the stack algorithm, the
number of branch metrics is averaged on a high number of blocks of K users. As
stated in Section 5 the numerical complexity of the stack algorithm decreases
rapidly as the SNR increases. For Eb/No 2': 5dB (BER:S 8.0 10- 3 ) the
numerical complexity of the stack algorithm is very close to its lower bound
(2K branch metrics computed).
The numerical complexity of the stack algorithm is then compared to that
of the feedback decoding algorithm, for window sizes L = 3, L = 2 and
L =1. Note that in the case L = 1 the number of branch metrics computed
is (K - L + 2)2£ - 2 = 2I<. In the case of the feedback decoding algorithm,
when the window is shortened by one unit (L -t L - 1), the number of branch
metrics computed is divided by about 2.
Figure 5 shows the BER obtained with the stack algorithm, the feedback
algorithm (L = 1,2,3), the decorrelator, as well as the single user bound.
As one can see from Figure 5 the stack algorithm (exact ML) performs always
better than the feedback decoding algorithm (approximate ML). What is more,
the feedback decoding algorithm performs always better than the decorrelator,
even when L = 1.
191
70~------------'------r~~~~~==========~
"0
Ql
:J
a.60
E
8
~
",S 50
Ql
E
..c
o
c
t1l40
15
'0
lii
~30
::J
c
Ql
Ol
~ 20
Ql ~~--~~--~~--~~--~~--~~--~~~~~
~
1~5L--------------L--------------L-------------~10
Eb/NO (dB)
The gain obtained with the stack algorithm, by comparison to the feedback
decoding algorithm, is all the more important as the SNR is high. At high
SNR the stack algorithm achieves the single user bound (Figure 5) with the
same computational complexity as that of the feedback algorithm for L = 1
(Figure 4). Compared to the decorrelator a gain of about 3.6 dB is achieved
by the stack algorithm at BER= 10- 5 .
7 Conclusions
In this paper, we have proposed a new approach to perform ML detection
for symbol vectors in DS-CDMA and MIMO systems. It is based on a QR
decomposition of the signatures matrix, or of the channel mixing matrix, fol-
lowed by an exact or approximate shortest path detection in a tree diagram.
In particular, using the stack and the feedback tree search algorithms yield
very attractive ML detectors both in terms of computational complexity and
performance.
lO-'L----'-----'-----'-----'----'----'--'-----'
o 4 6 8 10 12 14
Eb/NO (dB)
References
[1] G. Foschini. Layered space-time architecture for wireless communication in a
fading environment when using multi-element antennas. Bell Labs Technical
Journal, pages 41-59, 1996.
[2] R. V. G.D. Golden, G.J. Foschlni and P. Wolniansky. Detection algorithm and
initial laboratory results using v-blast space-time communication architecture.
Electronic Letters, 35:14-16, January 1999.
[3] G. Golub and C. Loan. Matrix Computations. The Johns Hopkins University
Press, Baltimore, Maryland, 1984.
[4] J. Heller. Advances in Communication Systems, volume 4, chapter Feedback
decoding of convolutional codes. A.J. Viterbi, New York, academic edition,
1975.
[5] F. Jelinek. Fast sequential decoding algorithm using a stack. IBM Jour. Res.
Dev., 13:675-685, November 1969.
[6] M. Junti. Multiuser demodulation for DS-CDMA systems in fading channels.
PhD thesis, Gulu University, Department of Electrical Engineering, Gulu, Fin-
land, 1998.
[7] I. Medvedev and V. Tarokh. A channel shortening multiuser detector for ds-
cdma systems. In VTC'Ol, Rhodes, April 2001.
[8] J. G. Proakis. Digital Communication. New York McGraw-Hill, second edition,
1989.
193
INTRODUCTION
An iterative decoding algorithm ("Block Turbo Code (BTC) algorithm") for product
codes based on soft decoding and soft decision output of the component codes was
introduced by R. Pyndiah in 1994 [1][2]. It uses the concepts developed by C.
Berrou who proposed a technique to encode and decode a class of error correcting
codes, called "Turbo codes" CTC [3]. The BTC is based on the product code which
is a series concatenated coding scheme, introduced by Elias [4]. The information bits
are placed in a matrix. The rows of the matrix are encoded by a linear block code
and the columns by a second block code.
Numerous papers, [5] to [12], describe this algorithm, give its performance, attempt
to simplify it and propose architectures for implementing it in integrated circuits or
DSPs.
For a better understanding of this paper we recall the basic principle of product
codes and their properties. We give a brief description of the soft decoding
algorithm (Section 1) and the soft decision algorithm (Section 2). In Sections 3 and
4, we analyze the performance of BTCs when the input data are binary, which is the
case in several systems (optical transmission, data storage, networking, ... etc).
The concatenated code used for BTCs is the product code proposed by Elias in
1954. Let us consider the concatenation of two systematic linear block codes C\ and
CZ with parameters (nt.kt.bt) and (nz,kz,b z) where nj, kj, and bi stand for code length,
number of information bits and minimum Hamming distance respectively. The
product code is obtained by :
1) placing the k\.kz information bits in an array of k\ rows and kz columns;
2) coding the k t rows using code CZ;
3) coding the nz columns using ct.
195
X. Lagrance and B. labbari (eds.),
Multiaccess, Mobility and Teletraffic for Wireless Communications, Volume 6, 195-204.
© 2002 Kluwer Academic Publishers.
196
------ n2 --------1.~
I •
k 2
Lhecks
kl Information symbols on
n1 rows
1
j Checks on columns
LI.eCI<$
on
checks
The parameters of the product code are given by: n=nl.n2, k=kl.k2 and <5=<5 1.<52 • The
code rate is given by R=R 1.R2 where Ri is the code rate of code C i. The fact that the
minimum Hamming distance of the product code is the product of the minimum
Hamming distance of the elementary codes gives product codes a significant
advantage over parallel concatenation [6]. This is a direct result of a property of
product codes, which states that all the rows are code words of C 2 and all the
columns are code words of C I.
The elementary code used is the Bose-Chauduri-Hocquenghem (BCH) code which
is a systematic cyclic code, able to correct at least t errors in a block of n symbols.
The minimum distance dmin is at least equal to 2t+ 1. Thus, it is possible to add a
parity bit in order to increase the minimum Hamming distance and improve the
coding gain. These particular codes are named extended BCH codes. The concept of
product codes is a simple and relatively efficient method to construct powerful codes
(that is, having a large minimum Hamming distance dmin ) using simpler linear block
codes.
The iterative turbo decoding process can be achieved by cascading several
elementary decoders illustrated in Fig.2, where k represents the current half-
iteration, and
1. [R] is the received vector,
2. [W(k)] is a vector that contains the extrinsic information (which is the
difference between the output information and the input information) given by
the previous decoder concerning the reliability of the decoded bit,
3. [R'(k)]=[R]+a(k).[W(k-l)],
4. a(k) and ~(k) are constants determined by simulations.
197
[R] [R]
D::layline
The basic element of a turbo decoder is the SISO decoder used for decoding the rows
and columns of the product code. It is a modified Chase algorithm [13] which starts
by computing the maximum-likelihood (ML) code word using the log-likelihood
ratio (LLR) of the bits at the input of the SISO decoder. For each bit of the ML code
word, it then computes the log-likelihood ratio which is the soft output of the
decoder [2].
2. RELIABILITY COMPUTING
Let us consider the transmission of binary elements {O, 1} coded by a linear block
code C with parameters (n,k,J) on a Gaussian channel using binary symbols {-1,+1}.
We shall consider the following mapping of the symbols: 0-->-1 and 1-->+1. The
output of the Gaussian channel R=(r[, ... ,rj, ... ,rn) for a given transmitted codeword
E=(e[, ... ,ej, ... ,en) is given by:
R=E+G (1)
where components gl of G=(g[, ... ,gj, ... ,gn) are AWGN (Additive White Gaussian
Noise) samples of standard deviation cr.
N~}~l~+ 1=I/oti
f~~~) (J
w
where:
r0 if c7 1Ul = ci 1(j) (7)
PI = ~ll if c+ 1(j)
I
oF- c- 1(j)
I
If we assume that (J is constant, we can normalize N(dj ) with respect to the constant
2/(J2 and we obtain the following equation:
(8)
with:
n 1(j) (9)
wi = "[.rIC: PI
1=lhi
The normalized LLR r j is taken as the soft output of the decoder. The term Wj is a
correction term applied to the input data and is called extrinsic information. The
extrinsic information is a random variable with a Gaussian distribution since it is a
linear combination of identically distributed random variables. Furthermore, it is
uncorrelated with the input data rj. Like for CTC, the extrinsic information plays a
very important role in the iterative decoding of product codes.
Computing the reliability of decision 0 at the output of the soft-input decoder
requires two code words C+l(/J and C1(J), see (5). Obviously D is one of these two
code words and we must fmd the second one which we shall call C. C can be viewed
as a competing code word of D at minimum Euclidean distance from R with Cj 0.
Given code words C and D, it can be shown that the soft output is given by the
following equation:
r' .=(IR-Cl 2 -IR-Dl 2
) 4
)d.
J
(10)
In the event where code word C is not found, we use the following equation:
rj=~ ..dj with ~ o. (11)
This sub-optimal solution uses reliability ~ (a constant for the current half-iteration)
at many positions}. As a refmement, we propose to estimate N(0) when there is no
competitor by:
x and y being chosen among the least reliable binary symbols of [ R'm] [14] [11].
199
If we compare the results for fixed and variable ~ (12), the experimental result is
much closer to the theoretical curve (gap of 0.1 dB to 0.2dB).
In order to further improve the estimation of ~ we can adopt the following strategy:
3. IMPLEMENTATION OF BTC
The "optimal" turbo decoding algorithm of linear block codes follows the following
steps:
1. Search for the least reliable binary symbols of [R']; their positions are called I],
I 2 .... I m ,
2. Generate test sequences [TQ] that are a combination of elementary test vectors
[TQy having "1" in position Ij and "0" elsewhere,
3. For each test word [TQ], compute [ZQ]:
[ZQ]=[TQ]ffi sign of [R'],
4. Decode [ZQ] by the algebraic algorithm (result [CQ]),
5. For each vector [C Q], compute the square Euclidean distance between [R] and
[C Q],
6. Select code word [Cd] at minimum distance from [R']; then [D]=[C d] is the
result of binary decoding,
7. Compute reliability Fj for each element dj of [D]; this involves searching for a
code word which has a minimal square Euclidean distance from [R'] (called a
competitor) with C/"*C/,
8. Compute extrinsic information:
d ' d
Wr[FrCj Rj ]Cj .
The implementation of this algorithm has shown that the complexity depends mainly
on the computation of the reliability at the output of the decoder. It is possible,
without any significant degradation, to decrease this complexity by reducing the
number of competitors. When using only one competitor, the loss is 0.12dB but the
complexity for this function is divided by 10. We show that the gap with theoretical
reference [14] is small, lower than 0.2dB at 10-6 • This very small value shows that
the algorithm allows the theoretical limit to be reached. The complexity of the
elementary decoder is very low: fewer than 10,000 gates are necessary to implement
the BCH decoder in the BTC solution.
200
ffi
III
I,OO&OS t--,--=+=-=-c*",=t~rlr=-~-=-_-=-_-="_=""'
__="",_,""_,""",_,...,,-1_
1,00&07
2 2,5 3 3,5
BllNOindB
Fig. 3 gives the Bit Error Rate (BER) as a function of the Signal Noise Ratio (SNR)
when the modulation used is QPSK modulation [9][11]. For this product code
BCH(32,26,4)2, the number of competitors varies. Sequential decoding with 6
iterations is used with the Chase algorithm using 16 test patterns. We measure the
performance of the BTC when using 5 bits for the linear quantization of the data
[R] and [W]. The code rate is about 0.66. These results show that BTCs are very
attractive for a wide range of applications because they offer the best trade-off
between performance and complexity.
Initially, turbo codes were designed for systems where soft data inputs to the
decoder were available. However there are many applications where the inputs are
binary. This is true for optical transmission systems and high speed data storage
applications. In classical forward error correction there is a loss of 2-3 dB in the
coding gain when using binary input instead of soft inputs. This is due to the fact
201
that using binary inputs, optimum decoding consists in finding a code word at
minimum Hamming distance which yields a coding gain of
G a 10.log (R(t+ 1))
where R is the code rate and tthe error correction capability given by t =r (<5-1 )/ 21
With soft inputs, optimum decoding consists in finding a code word at minimum
Euclidean distance which yields a gain of
G a 1O.log (R <5).
For product codes with <5=36 or <5=16 the difference in coding gain is ~G 3 dB.
In this paper we consider the degradation in coding gain as we decrease the number
of quantization bits down to one with BTC.
Figure 4 shows the behaviour of BTCs, as we change the number of quantization
levels. The code used in this diagram is the product code BCH(32,26,4)2. The
simulations are given for 2, 3, 4, 7, 8 and 31 levels. The first noteworthy result is
that, between binary quantization (2 levels) and 31 levels, the difference is around
2dB at BER = 10-6 , which is less than 3 dB predicted by theory. If we introduce
erasures in the input data (3 levels), an improvement of O.5dB is obtained at BER=
10-6 . The impact of quantization levels depends on factors such as the error
correcting capability or length of the codes. When the error correcting capability
increases, the difference between 31 levels and 2 levels decreases. For example, for
the product code BCH(32,21 ,6)2 the loss between 31 levels and 2 levels is 1.5 dB at
BER=1O-6 instead of3 dB.
0::
w 1,OOE-04
III
1,OOE-05
1,OOE-06
1,OOE-07
Eb/NO in dB
Figure 5 shows that for the product code BCH(128,113,6)2 the difference between
31 levels and 2 levels is around IdB at BER=1O·6 , and is only O.8dB when the
product code is BCH(256,239,6)2, which is a very long code.
-= - -
- --
- ~ --
-- . ~
-- -- --
-- --. ---
-- -- ~
1,00E-07 -+-
_
!256,239) 31 levels
256,239) 2 levels ~ ~ ~
--
- . - 128,113) 31 levels :::
___ 128,113) 2 levels - - -
1,00&08 .Ll::==~:::I::i::ci:=ii:r..u...u..Ll
2,8 3,1 3,4 3,7 4 4,3 4,6
Eb'NOindB
Fig 5: Performance for the product code BCH(J28,113,6/ and BCH(256,239,6/ at
iteration 8 as a jUnction of the number ofquantization levels.
Another way to evaluate the performance of the block turbo code is to compare the
bit error ratio BER(OUT) at the output of the decoder with the BER(lN) at the input
(see Fig.6).
203
BER(OUT)
10° ,------------------,
BER(IN)
Fig 6: Bit error ratio BER(OUT) as a function of BER(IN) for different BTCs.
We observe the very rapid variation of the output BER with input BER as we
increase minimum distance of component codes and their code length. For product
code BCH(256,239,6)2, the extrapolated BER(OUT) is very small «10- 15 ) for an
input BER < 1%.
5. CONCLUSION
BTC have been extensively investigated in the last eight years and their main
properties are:
-large minimum distance (16,24,36".),
-very high code rate (up to 0.96),
-low complexity « 10.000 gates),
-very high data rate (>50 Mbps).
In this paper we have investigated the effect of input data quantization on the
performance of BTC. Simulation results given here show that the loss in
performance when feeding binary data input instead of soft data input to a BTC
decreases as the code gain increases. In the case of BCH(256,239,6)2 the loss is of
0.8dB instead of the theoretical 3.0 dB loss. Although the theoretical justification is
not yet established, this result is extremely important for applications where only
binary data is available at the decoder input and should open the way to considerable
improvement of quality of service on heterogeneous networks.
REFERENCES
[1] R. Pyndiah, A. Glavieux, A. Picart and S. Jacq, "Near optimum decoding of
products codes," in proc. of IEEE GLOBECOM '94 Conference, vol. 113, Nov.-Dec.
1994, pp. 339-343.
204
[2] R. Pyndiah, "Near optimum decoding of product codes : Block Turbo Codes,"
IEEE Trans. on Comm., vol 46, nO 8, August 1998, pp. 1003-1010.
[3] C. Berrou, A. Glavieux and P. Thitirnajshima, "Near Shannon limit error-
correcting coding and decoding: Turbo-codes," IEEE Int. Con! on Comm.ICC'93,
vol 2/3, May 1993, pp. 1064-1071.
[4] P. Elias, "Error-free coding," IRE Trans. on Inf. Theory, vol. IT-4, Sept. 1954,
pp.29-37.
[5] P. Adde, R. Pyndiah, O. Raoul, "Performance and complexity of block turbo
decoder circuits," Third International Conference on Electronics, Circuits and
System ICECS'96, Rodos, Greece, 13-16 October 1996 -, pp 172-175.
[6] R. Pyndiah, "Iterative decoding of product codes: Block Turbo Codes," in proc.
of IEEE International Symposium on Turbo Codes & Related topics, Sep. 1997,
pp.71-79.
[7] P. Adde, R. Pyndiah, O. Raoul and J.R. Inisan, "Block turbo decoder design,"
International Symposium on turbo codes and related topics, Brest, Sept. 1997,
pp.166-169.
[8] A. Goalic and R. Pyndiah, "Real time turbo decoding of product codes on a
digital signal processor," in proc. of IEEE International Symposium on turbo codes
and related topics, Brest, Sept. 1997,pp. 267-270.
[9] S. Kerouedan, P. Adde et P. Ferry, "Comparaison performances/complexite de
decodeurs de codes BCH utilises en turbo decodage", GRETSI'99, Vannes, 13-17
Sept. 1999, pp 67-70.
[10] S. Kerouedan and P. Adde, "Implementation of a block turbo decoder in a
single chip", 2nd International Symposium on turbo codes and related topics, Brest,
Sept. 2000, pp. 243-246.
[11] S. Kerouedan, P. Adde and R. Pyndiah, "How we implemented block turbo
codes?", Annales des Telecommunications, tome 56, juilletlaout 2001, pp. 447-454.
[12] S. Kerouedan and P. Adde, "Block turbo codes: Towards implementation", 8th
International Conference on Electronics, Circuits and System ICECS'200J, Malta,
September 5-7, pp. 1219-1222.
[13] D. Chase, "A class of algorithms for decoding block codes with channel
measurement information," IEEE Trans. Inform. Theory, vol IT-18, Jan. 1972, pp.
170-182.
[14] P. Adde and R. Pyndiah, "Recent simplifications and improvements of Block
Turbo Codes", 2nd International Symposium on turbo codes and related topics,
Brest, Sept. 2000, pp. 133-136.
Gallager Codes for Asynchronous Multiple Access.
Introd uction
205
X. Lagrance and B. Jabbari (eds.),
Multiaccess, Mobility and Teletraffic for Wireless Communications. Volume 6, 205-220.
© 2002 Kluwer Academic Publishers.
206
1. Channel model
sdl-l]
..
i
symbol period
..
s~dl-l]
...
. ....
1- TNu
N-1
factorization theorem (Lehman, 1959), { y(1) [1], y(2)[I], ... ,y(Nu ) [1] } /=0
are sufficient statistics to estimate the transmitted messages. This im-
plies that the channel output {y(t)} enters in the computation of the
posterior probability of each message only through y(n) [1]. See (Verd ti,
1989a) for more details. The basic discrete time AWGN asynchronous
multiple access channel with total input power constraint, rectangular
pulse and noise variance N o/2 can also be modeled by:
Nu Nu
Y= L anMT,nsn + n = L MT,nXn + n = MTx + n = z + n (1)
n=l n=l
(2)
where b n is a vector of K information bits. We suppose in the sequel
that the frame length N, the code rate R are equal for all users, we are
frame synchronous and one trailing zero is added at the beginning of
each frame to ensure the causality of the transmission.
The aim of this section is to recall the performance limits of the AWGN
asynchronous MAC for BPSK sources (E {-I, I} ). This analysis will be
useful to evaluate the performance of the proposed method. Under the
assumptions of the previous section, we adopt the following definition
according to (Gilhousen et al., 1991; Verdu and Shamai, 1999):
(3)
Before we derive the capacity, i.e. the maximal spectral efficiency, of the
asynchronous AWGN-MAC with BPSK sources, we recall the formula of
the capacity of the asynchronous AWGN-MAC with Gaussian sources.
Under the global power constraint, i.e. the average power per symbol
209
where I denotes the identity matrix, 1.1 the determinant operator and
N the code block length.
With two users, it has been shown in (Verdu, 1989a) p.737 that the
worst case offset between the signals is zero, i.e. in which case the channel
is symbol synchronous. The most favorable case occurs when the symbol
offset is equal to half the symbol period. We observe the same results
with two BPSK sources.
A2) all users are equal power: IE [s;[I]] = 1, \In = 1, ... ,Nu
8rr=~~~~~~~~----~r------'
+ BPSK Nu=2 - a~=OT (sync.)
7 - - BPSK Nu=2 - a~=O.25T
o
-1
-~L-------~------~------~~------~2
Figure 2. Ebj No vs. capacity: Gaussian sources, BPSK 2 users with equal power
(asynchronous case: AT = O.25T and O.50T, synchronous case: AT = OT).
R and the same power. In that case, the new code rate R'(R) is given
by (6).
Figure 3. BER vs Eb/ No for different rates (R = t, t) - Equal power BPSK async.
2 user (b.T = OT, O.25T, O.50T).
These block codes have been proposed by Gallager in 1963, together with
a stochastic decoding algorithm which is very close to belief propagation.
Mc Kay & al. have rediscovered and extended LDPC Gallager codes
recently (MacKay, 1999) and have shown that Gallager codes can be
easily decoded with iterations of belief propagation on their factor graph
(d. figure 4).
First, we describe the functional nodes (black square). Since it is the
single user case, we have: x[/] = s[/]. Each channel node calculates the
conditional probability densities:
p(x[/]Jy[l]) ex ~exp
27r(J2
(-(y[l] - x[/])2/2(J2) (7)
213
Each parity-check node indicates that the set Qk of the codeword bits
{s[l]} E Qk to which the parity-check is connected have even parity:
channel
output
channel
node
codeword
parity-check
node
!
In the multiple user case, we rewrite the model described by (1):
y(Nu ) [I] ~
214
Then, each codeword xn[l] is connected to both variables z[l-l] and z[I].
Figure 5 shows the factor graph for a joint asynchronous multiple user
system using Gallager codes. The fading coefficients an are supposed to
be perfectly known at the receiver.
As in the single user case, each channel node calculates the conditional
probability densities for the channel:
In the multiple user case, z[l] is described by (10). The variable z[l] has
2Nu components and then 22Nu possible states.
We define, as the "spine-check" node, the functional node to which
z[I], z[1 - 1] and Sn[l] , iiI :S n :S Nu are connected. Using (10), this
functional node is described by:
Nu
L anMT,nsn[l] = z[l] (12)
n=l
Figure 5. A factor Graph for a joint asynchronous multiple user decoding algorithm
using Gallager codes C(N, M)
3.3.2. Decoding
The major limitation of such joint multiple user detection algorithm is its
exponential complexity in the number of users. Fortunately, there exist
various means in order to reduce this complexity. A well-known result
in the graph theory, see for instance (Frey, 2000) is that the exponential
complexity at the spinal node can be reduced to a polynomial complex-
ity o(N~) with no loss in performance. In most of cases, this remains
too complex. Several suboptimal strategies are possible. For example,
216
4. Simulation results
1O~'0'---0"-.5----'--'------'--'-2_----:'-c_--'-_----,-'------'
EblNO [dB]
10-'oL--:'-:----'---:":---:---:'-::----:'-----::'-:-----'
EblN~[dBl
Conclusion
bounds with respect to the code rate R, the number of users N u , the sig-
nal to noise ratio Eb/NO and the BER. Note that these bounds generally
are not equal to the single user bounds. The performance of our system
are close to the optimum achievable bound (2dB or less at BER= Ie - 3
for a block length N = 2000) for a system load RNu equal to 1 and 0.5
(R = 1/2 and 1/4). The major problem is still the complexity of the
algorithm exponential in the number of users. As described in section
3.3.2, a suboptimal method combining multistage and iterative detection
is possible.
References
Berrou, C., A. Glavieux, and P. Thitimajshima: 1993, 'Near Shannon Limit Error-
Correcting Coding and Decoding: Turbo Codes'. In: ICC. pp. 1064-1070.
Chung, S.-Y., T. Richardson, and R Urbanke: 2001, 'Analysis of sum-product de-
coding of low-density parity-check codes using a Gaussian approximation'. IEEE
Trans. Inform. Theory 47, 657-670.
Cover, T. and J. Thomas: 1991, Elements of information theory. New york: Wiley
edition.
de Baynast, A. and D. Declercq: 2002, 'Gallager codes for multiple access'. accepted
to IEEE Symposium on Information Theory.
Duff, I. S., A. M. Erisman, and J. K. Reid: 1986, Direct methods for sparse matrices.
Clarendon Press; Oxford University Press, oxford: New-York edition.
Frey, B.: 2000, Graphical Models for Machine Learning and Digital Communication.
Cambridge, Massachussets: The MIT Press.
Gallager, R: 1962, 'Low-Density Parity-Check codes'. IRE Transactions on Infor-
mation Theory.
Gilhousen, K. S., I. Jacobs, R Padovani, A. Viterbi, 1. Weaver, and C. Wheatley:
1991, 'On the Capacity of a Cellular CDMA System'. IEEE Trans. on Vehicular
Technology 40(2).
Johansson, A. and A. Svensson: 1995, 'Multi-stage interference cancellation in multi-
rate DS/CDMA systems'. In: PIMRC'95. Toronto, Canada.
Kschischang, F., B. Frey, and 1. H.-A.: 2001, 'Factor graphs and the sum-product
algorithm'. IEEE Trans. Inform. Theory 47(2), 498-519.
Lehman, E.: 1959, Testing Statistical Hypotheses. New york: Wiley edition.
MacKay, D.: 1999, 'Good Error-Correcting Codes Based on Very Sparse Matrices'.
IEEE Transactions on Information Theory 45.
MacKay, D., S. Wilson, and M. Davey: 1998, 'Comparison of constructions of irregular
Gallager codes'.
Patel, P. and J. Holtzman: 1994, 'Analysis of a simple successive interference can-
cellation scheme in a DS/CDMA system'. IEEE Journal on Selected Areas in
Communications 12(5).
Scaglione, A., G. Giannakis, and S. Barbarossa: 2000, 'Lagrange/Vandermonde MUI
Eliminating User Codes for Quasi-Synchronous CDMA in Unknown Multipath'.
IEEE Transactions on Signal Processing 48(7), 2057-2073.
220
1. INTRODUCTION
Communication systems use increasingly complicated channel coding and
modulation techniques to gain a higher throughput with lower power
consumption. Designs such as adaptive modulation and coding (AMC) and
hybrid automatic repeat request (HARQ) are likely to be used in 3G and 4G
systems. A set of operating modes must be determined, each with a different
combination of parameters including the frame size, the code rate, and the
modulation order. It is desirable to have a quick and systematic method to
find the correct combination.
One method is to use bounding techniques. With the capacity-approaching
channel coding methods, such as turbo codes [1] and low-density parity-
check codes, it is possible to find a theoretic bound which is closely
approached by a well-designed code [2]. The bound includes the effect of
finite block size and realistic modulation, where the modulation is
considered a part of a composite channel. The bounds can quickly provide
valuable information on the potential performance of the system without
requiring extensive simulations. Methods making use of the bounds have
been applied to improve the design of a HARQ system for 1xEV-DV [3].
221
X. Lagrance and B. labbari (eds.),
Multiaccess, Mobility and Teletraffic for Wireless Communications, Volume 6, 221-238.
© 2002 Kluwer Academic Publishers.
222 Yufei W Blankensip and Brian K. Classon
rate R M = 2 q constellation
k info bits
estimated
info bits
where q(Xi) is the probability distribution function of the channel input Xi. Let
Cw and C2D represent the capacity limits for anyone dimensional (1-D) and
Bounding Techniquesfor the Design of Channel Coding and Modulation Systems 223
Cw = .!..10 g (1 + 2Es)
2 (info bits/symbol). (2)
2 No
When the capacity limit is achieved, the signaling rate is r = Cw . Thus, the
minimum bit SNR requirement is
(~E)w 22ClD -1
2Cw
(3)
For 2-D signaling with equal noise power on each dimension, the capacity of
the 2-D AWGN channel is
(4)
where Es is the average energy of the vector symbol and equal to twice the
symbol energy of I-D signaling. Considering Es = rEb = C2nEb, the minimum
Eb/NO is
(5)
(6)
where
224 Yulei W Blankensip and Brian K. Classon
(7)
and the average power per symbol Es = _1_ I (Ai 2 Bi 2) is used to normalize
M i~O
+
the symbol energy in (7). The function D(u,v) accounts for the location of the
constellation points. For instance, an M-PSK constellation has Es = 1, and
can be defined as Aj = cos 27rj/M and B j = sin 27rj/M .
Figure 2 shows the composite channel capacity found by (6) for QPSK,
8-PSK, 16-PSK, 16-QAM, 32-QAM, and 64-QAM. The 2-D capacity limit
C2D is also plotted for comparison. Each modulation uses the conventional
constellation structure, i.e., PSK symbols are evenly spaced on a circle, and
the QAM symbols are evenly spaced on a rectangular grid.
Some design guidelines can be derived from Figure 2. The maximum
number of information bits that can be transmitted per channel symbol is
determined by the lowest amount of redundancy introduced by channel
coding (i.e., highest code rate). Since when the capacity is reached,
r = C = Rq, R = C/q, thus the minimum Eb/NO is actually a function of the
channel coding rate R. As expected, each modulation with M (= 2q)
constellation points achieves the maximum of q information bits/symbol, or
code rate R = 1, as Eb/NO becomes large (Eb/NO> 16 dB for all under
discussion). As M increases, the number of information bits per pulse-
amplitude modulation (PAM) symbol approaches Cw , while that of PSK and
QAM approaches C2D • The gap between the C2D and the modulations in
Figure 2 is due to the nonideality of the modulation techniques.
Bounding Techniques for the Design of Channel Coding and Modulation Systems 225
13PSK •
QPSK :
BPSK :
4 6 8 10 12 16
Minimum E"INo (dB)
Figure 2. The composite channel capacity (info bits/symbol) vs. minimum E,)No (dB) for
modulations BPSK, QPSK, 8-PSK, 16-PSK, 16-QAM, 32-QAM, and 64-QAM. Unquantized
detection and a static channel are assumed. The 2-D capacity limit C2D is plotted for reference.
Within the same modulation category (PSK or QAM), Figure 2 shows that
the bandwidth efficiency, measured by signaling rate r (= C), is always larger
when M is higher. However, in general, high M implies high signal
processing complexity. Thus, for the bandwidth efficiency region where
there is little difference between higher- and lower-order modulation, the
lower-order modulation should be used. For example, both R = liz 64-QAM
and R = % 16-QAM provide r = 3 info bits/symbol. Since there is only
0.35 dB difference in performance and 16-QAM requires less signal
processing, it may be more appropriate to use R = % 16-QAM. Figure 2
suggests that for the same modulation categoryl, the system can use a 2Q-ary
modulation for a code rate R < 0.8 at the expense of less than 0.5 dB instead
of a 2Q+ l_ary modulation with a code rate Rq/(q+ 1). In other words, higher
code rate with lower order modulation may be more appropriate than lower
code rate with higher order modulation. However, as R gets closer to 1
(C~q), the curve of a given modulation flattens, and the increase of the
minimum E,JNo brings quickly diminishing returns in bandwidth efficiency
for the increased code rate. It is better then to use a higher level modulation.
1 The rule does not apply to BPSK and QPSK since the dimensionality changes between
BPSK and QPSK.
226 Yufei W. Blankensip and Brian K. Classon
1.... .. ....................... .
....... :
0.5 ..
...... ,.......
-0.5 ...... , ..........,.................. .
-0.5
-1
-1 0 1 -1 0 1
(e) d ~in/average energy =0.845 (d) d~in/average energy =0.928
Figure 3. Four constellations of 8-QAM and their figures of merit. (a) 8-PSK, (b) rectangular
8-QAM, (c) optimal rectangular 8-QAM, (d) hexagonaI8-QAM.
Bounding Techniquesfor the Design of Channel Coding and Modulation Systems 227
Figure 4. The composite channel capacity (info bits/symbol) vs. minimum EJNo (dB) for the
four 8-QAM constellations in Figure 3 with AWGN channel.
In Figure 3, four typical constellations of 8-QAM are shown with unity
distance between adjacent points. Constellation (a) is 8-PSK and (d) is
constructed on the hexagonal lattice2 . Constellations (b) and (c) are
constructed on a rectangular lattice, with (c) being the best rectangular
design [9].
From the d~n/Es shown underneath their graphs in Figure 3, it can be
predicted tentatively that 8-PSK has the worst power efficiency, and the
hexagonal 8---QAM has the best. To test their performance, their composite
channel capacities vs. minimum EblNo are calculated and plotted in Figure 4.
The curves do show that the hexagonal design outperforms the other three.
The difference between the four designs is mainly in the high code-rate
region 0.6 < R < 1 (i.e., 1.8 < C < 3). The hexagonal 8-QAM shows as much
as 1.2 dB advantage over 8-PSK. However, there is very little difference
between the hexagonal 8-QAM and the optimal rectangular design.
The same study was also carried out on 16-QAM constellations. The
optimal 16-QAM constellation is about 0.2 dB better than the conventional
16-QAM. Thus there is little advantage to use the optimal constellation in
terms of the composite channel capacity when a good rectangular
constellation can be used. Considering the extra complexity required to
implement a mathematically optimal constellation, the small gain may not
justify abandoning the conventional rectangular QAM constellations.
2 The shape refers to the lattice, or the regular array of points, from which the signal points are
selected [6]. The densest two-dimensional lattice is the hexagonal lattice [8].
228 Yulei W. Blankensip and Brian K. Classon
3 The set of sequences at distance d or less from a codeword can be interpreted as a sphere of
radius d around the codeword. Sphere packed codes are the set of codewords such that the set
of spheres of radius d around the different codewords exhaust the space of binary n-tuples and
intersect each other only on the outer shells of radius d. Shannon's perfect code for the
continuous-input channel requires that the entire continuum of n-dimensional Euclidean space,
not just the discrete points represented by binary n-vectors, be filled up by the nonintersecting
cones. This is much stricter than the requirements of a normal sphere-packed code.
Bounding Techniques for the Design of Channel Coding and Modulation Systems 229
and
. 11
G smuexp -A + AGcose
2 )]"
[ (
4Shannon pointed out that the perfect cone partitioning, and thus the perfect spherical code, is
only possible for n = 1 or 2, if k > 1 [10].
230 Yufei W Blankensip and Brian K. Classon
While all three bounds converge as k ~ 00, both the random coding bound
and the channel capacity bound have little meaning for low k.
Figure 6 shows that the word error probability Pw, not the signaling rate r,
affects the shape of the curve. Regardless of rand P w, the sphere packing
bound approaches (EbINo)w as k ~ 00. For a given Pw, curves of all rates
Bounding Techniques for the Design of Channel Coding and Modulation Systems 231
approach their respective capacity limit similarly. The difference between the
asymptotic bound and the exact bound in the transition region is more
pronounced for higher Pw than for lower Pw. For the code rates examined, the
asymptotic curves merge to the exact curves around k = 1 ~ 100, with
smaller k for lower P w, and larger k for higher P w. Thus, for the block sizes
°
k> 100, the asymptotic expressions of the sphere packing bound can be used
instead of the computationally prohibitive exact expressions.
The discussion above of the sphere packing bound is limited to I-D
signaling. However, the parameters of ND-dimensional signaling can be
normalized so that the sphere packing bound can be calculated for a system
with ND-dimensional modulation as well. As stated, the derivation of the
sphere packing bound is based on each codeword of n symbols being seen as
a point in an n-dimensional Euclidean space. For any ND-dimensional
modulation, a frame of n symbols is a point in an ND x n-dimensional
Euclidean space. Thus finding the optimal performance of a code with k info
bits/frame, r info bits/symbol, and ND-dimensional modulation is equivalent
to finding the sphere packing bound with the normalized parameters: info
frame size k, codeword size n' = (k/r) x N D , and per-dimensional signaling
rate r' = r/ND = Rq/ND where r can be greater than 1. When applied to the
system in Figure 1 with a 2-D modulation (such as QPSK and 16-QAM), the
lower bound Pw,spb is found by using the set of parameters (k, n' = 2k/R/q,
r' = Rq/2) in the sphere packing equations in place of (k, n = k/R/q, Rq). This
lower bound applies to any channel coding with realistic modulation.
(a)Pw =IO- 1
Figure 6. The required Et/No derived from (i) channel capacity limit C w , (ii) Shannon's sphere
packing bound, and (iii) Shannon's random coding bound for a range of information block
sizes k and signaling rates r E {1I3, 112,2/3,3/4, 7/S} to achieve word error probabilities of
P w E {lO-I, 10-2, 10-3 , 1O-4 } in a continuous-input AWGN channel. For the sphere packing
bound and the random coding bound, exact expressions are used for small k while asymptotic
expressions are used for large k. The transition region is drawn using polynomial curve fitting.
-O'~O~,-=~.E=Et::±[jl0~'~~::!::h:b:::kWl0'
Intannatian Block Size (bits)
Figure 7. The adjusted sphere packing bound ofBPSK and Gallager's upper bound of binary-
input AWGN channel.
Code rate
Figure 8. Modulation imperfection LlMOD(R) in dB for QPSK, 8-PSK, 16-QAM, and 64-QAM.
For a given frame size, code rate R, and 2Q-modulation, the adjusted
sphere packing bound is obtained by adding ~MOD(R) to the sphere packing
bound of the given frame size and signaling rate Rq. Figure 9 illustrates the
procedure of calculating the adjusted sphere packing bound using QPSK as
an example.
In Figure 10, the adjusted sphere packing bound is plotted against the
simulated performance of a turbo-coded orthogonal frequency division
multiplexing (OFDM) system for an AWGN channel. In this system, the
number of OFDM subcarriers is fixed, and each subcarrier employs the same
modulation scheme (e.g., QPSK, 16-QAM, 64-QAM). The channel coding
Bounding Techniques for the Design of Channel Coding and Modulation Systems 235
(a) Measure LlQPSK(R) (b) Shift the sphere packing bound up by ~psK(R)
Figure 9. A QPSK example is used to illustrate the procedure of obtaining the adjusted sphere
packing bound. LlQPSK(R) is obtained by measuring the gap between the composite capacity of
QPSK and the 2-D capacity limit. Then the adjusted sphere packing bound is obtained by
shifting the sphere packing bound by LlQPSK(R).
236 Yufei W Blankensip and Brian K. Classon
10 ...
750 subc~rl~rs :
F:;= 1O-"'i8PSK,-tlJ~o code
Code Rate
16
14
12
in
~10
~:
ill 8
Code Rate
5. CONCLUSION
Performance bounds based on channel capacity and Shannon's sphere
packing bounds are discussed with application to. coding/modulation system
design. Unlike other bounding techniques, these bounds strive to take the
specific modulation into account. The performance bounds discussed here
can be used in many system design applications. They have been used to
theoretically calculate the maximum throughput of a HARQ system with a
given set of coding/modulation parameters for lxEV-DV [3]. They have
also been used to validate the coding/modulation design of 3G and 4G
systems [13]. Although this paper only discusses AWGN channels, similar
analysis can be applied to fading channels.
Bounding Techniquesfor the Design of Channel Coding and Modulation Systems 237
References
Abstract The continuing evolution of exponential growth of the Internet market and grow-
ing demand of data services in fixed networks is also expected for mobile users.
Universal Mobile Telecommunications System (UMTS) as the 3rd generation
mobile system is designed to provide high bit data rates to the user. The mo-
bile Internet access will playa key role to ensure success of UMTS introduc-
tion. Typical Internet applications are running on a Transmission Control Proto-
col/Internet Protocol (TCPIIP) stack which should ensure a reliable end-to-end
communication in systems with limited quality of service guarantee. In contrast,
most protocols of radio access systems will grant traffic contracts. This raises
the question if the benefits ofTCP might be a drawback for the overall perfor-
mance of the radio interface with its own quality of service management mecha-
nisms. The aim of this paper is to examine the interaction between TCP and the
UMTS Radio Link Control (RLC) protocol. Of special concern is the interaction
of TCP retransmissions with the RLC Automatic Repeat Request (ARQ) mech-
anisms. Simulations have been performed to evaluate the performance of the
implemented protocol stack with and without TCP. Simulation results of quality
of service parameters like delay and packet throughput depict the performance
of mobile Internet access over UMTS.
Keywords: UMTS, UMTS Radio Access, Radio Interface Protocols, WWW, Quality ofSer-
vice, TCPIIP, RLC
1. Introduction
The aim of Universal Mobile Telecommunications System (UMTS) is the
provisioning of high bit rate data services to the mobile user. Hence, typi-
cal Internet applications like World Wide Web (WWW) browsing will migrate
from fixed access network systems to the mobile environment. Normally these
applications are running on a Transmission Control Protocol/Internet Protocol
(TCP/IP) protocol stack which guarantees a reliable end-to-end communica-
tion. Nevertheless, these protocols will not ensure quality of service in terms
239
X. Lagrance and B. fabbari (eds.),
Multiaccess, Mobility and Teletraffic for Wireless Communications, Volume 6, 239-250.
© 2002 Kluwer Academic Publishers.
240
of delay and throughput. In most cases the customer faces best-effort resource
management strategies in fixed networks. This might be applicable since over-
provisioning of bandwidth is feasible in fixed networks.
Due to the nature of radio links, the radio interface protocols have to cope
with limited bandwidth, higher delays and error rates. Therefore, the UMTS
radio interface protocols provide various and highly complex functions and
mechanisms to realize a reliable communication. Traffic contracts and quality
of service requirements are supported if desired.
This paper demonstrates the opportunities of the UMTS Radio Link Control
(RLC) protocol to support quality of service. RLC provides a reliable link with
ARQ mechanisms. Both, TCP and RLC protocol are using retransmissions to
recover lost data packets. Running an Internet application over UMTS rises
the question if it is necessary to use TCP. On the one hand side, protocol
overhead burdens the radio interface with its limited bandwidth. On the other
hand side, TCP retransmissions and time outs may interfere with the Automatic
Repeat Request (ARQ) of the RLC protocol. Another opportunity is running
a TCP application over an unacknowledged RLC connection. Since TCP is
not designed for radio access, it has to be shown how it will perform in an
environment with high error rates and delays.
The aim of this paper is to present performance results with the help of a
UMTS Radio Interface Simulator. The performance evaluation considers the
detailed UMTS protocol implementation as well as the traffic load generation
for WWW browsing sessions including TCP/IP modeling. Thus, it is feasible
to present the performance of a TCP application in an unreliable mobile envi-
ronment. The simulation results will show the performance in terms of quality
of service parameters a mobile user will experience while surfing the Internet.
Following this introduction, Sec. 2 gives a description of the traffic model.
Furthermore it presents the used simulation methodology and the simulation
environment. Sec. 3 and Sec. 4 illustrate the quality of service mechanisms of
the TCP and the RLC protocol. Performance evaluation, simulation scenarios
and discussion of simulation results is depicted in Sec. 5 and Sec. 6. The paper
concludes in Sec. 7 with a summary.
2. Simulation Environment
A WWW browsing session is a typical application which is running on the
TCP/IP protocol stack. A WWW traffic model is necessary for simulative
examinations of the performance of data services of mobile radio networks. A
typical WWW browsing session consists of a sequence of page requests. These
pages contain a number of objects with a dedicated object size each. During
a page request, several packets for each object may be generated which means
that the page request constitutes of a bursty sequence of packets. The bursty-
Ness during the page request is a characteristic feature of packet transmission.
After the page has entirely arrived at the terminal, the user is consuming certain
amount of time for studying the information. This time interval is called rea-
241
ding time. Tab. 1 gives an overview of the used WWW traffic model described
by parameters and their distributions.
Related documentation can be found in Frost and Melamed, 1994; Arlitt and
Williamson, 1995; Paxson, 1994. The main part of the later implementation is
based on the work of Arlitt and Williamson, 1995.
For this study the simulation tool UMTS Radio Inteiface Simulator (URIS)
was developed at the Chair of Communication Networks. This simulation en-
vironment is used to investigate, optimize and develop features of the radio
interface protocol stack. In addition, it offers the opportunity of capacity and
quality of service evaluation by simulations of various scenarios. The simu-
lator is a pure software solution in the programming language C++. The si-
mulation model is implemented with the help of a powerful C++ class library
which was developed by the Chair of Communication Networks and is called
the SDL Peiformance Evaluation Tool Class Library (SPEETCL) Steppler and
Lott, 1997. This generic, object oriented library is well suited for telecom-
munication network simulation purposes and can be used in event driven, bit
accurate simulation environments. The UMTS protocols at the radio interface
enhanced by a TCPIIP protocol stack were specified with the Specification and
Description Language (SDL). To generate an executable out of the SDL phrase
notation and the C++ library, a SDL2SPEETCL code generator is used.
The software architecture of the URIS simulator is shown in Fig. 1. Up
to now the simulator consists of various traffic generators and a traffic load
mixture unit which is used to adjust scenarios with desired load mixtures. The
physical channel module models the transmission of bursts in radio frames on
the radio interface. This includes discarding of erroneous bursts depending on
the error model.
The core of the simulator are the modules User Equipment (UE) and UMTS
Terrestrial Radio Access (UTRA) which are built formally similar. Each UE
and UTRA is implemented as an SDL system which contains the protocol im-
plementation of the layers. Fig. 2 gives an overview of the protocol structure
in a generic SDL system. The complex protocols like Medium Access Control
Physical Channel
(MAC), RLC, Packet Data Convergence Protocol (PDCP) and Radio Resource
Control (RRC) based on UMTS Release 4 and the TCP/IP and User Data-
gram Protocol (UDP)/IP are specified in SDL with object oriented methods.
For simulation purposes, the TCP/IP protocol stack can be easily bypassed by
switching the transport and the network layer to transparent transmission.
Usual simulator approaches model protocols and functions on the basis of
abstractions and simplifications. The aim of URIS is a detailed, bit accurate
implementation of the standardized protocols. This offers the opportunity to
determine the performance ofUMTS in a realistic manner.
• Flow control.
243
Traffic Generators
SOL System
!
ENV
I---- LLME Transport Layer
DC TCP UDP CS
I I
TCP UDP CS
I- RRC Network Layer
LLME IP AM UMTI
t !
LLME
Ipocpl
!
-1 Radio Link Control
Ilogical channels
5. Simulation Scenarios
The main parameters concerning Quality of Service (QoS), which are af-
fected by RLC and TCP/IP, are delay and throughput. Two measurements are
made during the simulation in both the RLC layer and the transport layer:
Packet Delay: time from sending a traffic load packet to the respective
layer until correct reception of the packet by the traffic load receiver,
2 Packet Throughput: size of the traffic load packet divided by the packet
delay. The throughput is given in bytes per second [byte/s).
i~
0.8 O.B
i
c
0.6 0.6
!
" 0.4 i" 0.4
~
0.2
aekel Eno( Rale = 0%
acket EITOf' Rate = 1%
acket EnOl" Rate'" 3%
JJ
----- ....
I a:-"
A
0.2
acket~~
0 0
0 0.5 1 1.5 2 0 20000 40000 60000 80000 100000
Downlink Delay [s] Downlink Throughput [byteJs]
(a) Traflic Load Packet Delay with RLC AM and (b) Traflic Load Packet Throughput with RLC
without TCP/IP AM and without TCP/IP
1 --
acket Error Rate =0%
acket Error Rate = 1%
acket Error Rate = 3%
acket Error Rate 5%=
i~
0.8 0.8
acket Error Rate =10%
i
~
0.8 0.6
! 1"
""
v
0.4 0.4
a:-
J
0.2
0%
1%
~ 0.2
3%
5%
10%
0 0
0 0.5 1.5 0 20000 40000 60000 80000 100000
Downlink Delay [5] Downlink Throughput (byte/5]
(c) Traflic Load Packet Delay with RLC AM and (d) Traffic Load Packet Throughput with RLC
TCP/IP AM and TCP/IP
I . "\
1 \ r\\· .
1
o 8' . . \...
acket Error
~acket
,ekel
Rate =
Error Rale
Erro< Rate
acket Error
--0%
1%
acket Error Rate = 3%
=
Rale - 5%
10%
1 0. 6 \
1
~ . ~,
0.4
R
it" 0.2
(e) Traflic Load Packet Delay with RLC UM and (I) Traffic Load Packet Throughput with RLC
TCP/IP UM and TCP/IP
Figure 3. Simulation Results for Traffic Load Packet Delay and Throughput
246
The cumulative distribution function of the measured packet delay and the
complementary cumulative distribution function of the measured packet through-
put have been calculated. Measurements have been made for increasing packet
error rates from 0% up to 10%. Tab. 2 shows the parameters of RLC and
TCP/IP used for these simulations.
6. Simulation Results
The results of the simulations are given in Fig. 3. As can be seen a mini-
mum delay is always present, caused by the constant transmission delay of
the physical layer. Packet error rates of 0% represent the ideally achievable
throughput. The throughput is not limited by the underlying dedicated channel
capacity since the offered load is always below 2 Mbitls.
Parameter Value
TTl Length 0.04 s
Dedicated Channel Capacity 260 kbyte/s
RLC TxWinSize 1024 PDUs
Max. No. ofPDU Retransmissions 40
Status Prohibit Timer 0.08 s
Poll Timer 1.0 s
Maximum TCP Segment Size 512 byte
Maximum TCP Send Window 16 kbyte
Min. TCP Retransmission Timeout 3s
Max. TCP Retransmission Timeout 64 s
Delayed Acknowledgment Not used
Selective Acknowledgment Not used
Header Compression Not used
Packet Error Rate 0%,1%,3%,5%,10%
247
ets have a delay of less than I s. The throughput is approximately 2/3 of the
throughput at 0% error rate. For a 5% packet errorrate, the Go-Back-N mecha-
nism is no longer able to provide sufficient data throughput. Due to the un-
necessary Go-Back-N retransmissions of correctly received packets even more
transmission errors are caused and congestion occurs within the RLC protocol.
To avoid unnecessary retransmissions and congestion a Selective-Reject-
ARQ (SR-ARQ) mechanism should be used. The SR-ARQ is currently under
study. It is expected that the performance will be better but a Go-Back-N ARQ
as a first approach is sufficient to study the general behavior.
7. Conclusion
This paper comprises simulative examinations of the performance ofWWW
surfing over the UMTS radio interface. The simulation results show that a
WWW session using TCP and the RLC UM is not feasible to satisfy a mobile
user. If packets are lost the TCP protocol infers that there must be congestion
in the network between the two peers. In consequence, a TCP retransmission
leads to a decrease of the send window size which results in less throughput.
This assumption is correct in wired environments but TCP is unable to handle
the unreliable radio link of mobile users due to the high error rates and TCP's
slow error handling. Very high delays and small throughput lead to unsatisfied
mobile users. Improvements may be archived by using the TCP SACK option
which allows the receiver to convey more information about its state to the
sender, thus increasing the efficiency of the retransmission strategy.
Running a WWW session over a standard TCP and the RLC AM will be
an ordinary scenario during the introduction ofUMTS. First UEs will rely on
standard application solutions which will be adopted from the fixed world.
Hence, TCP will run end-to-end including the radio interface. This paper
shows that this solution is applicable to satisfy the mobile user but the perfor-
mance suffers since TCP mechanisms will not efficiently use the guaranteed
QoS of the UMTS radio bearers in terms of delay and throughput. It has to
be mentioned that effects of the core network are not included in our simu-
lations. This is not only due to currently missing simulation capabilities but
also partly unknown structure of the core network. While it is hard to model
the behaviour of packets traveling along the internet, it can't even be assumed
that the entire core network is IP-based. Thus it can be concluded that the
overall performance will get worse since TCP as an end-to-end protocol will
experience additional delays and congestion in the core network. Using heav-
ily disturbed radio links the mobile user will face a reduced throughput and
longer waiting times for WWW page downloads than in fixed networks. This
249
relies on the weak TCP performance in a mobile environment and can not be
influenced by assignment of radio bearers with higher capacity.
The simulation results of a WWW session running without TCP/IP in RLC
AM show the best performance. As a result of all simulations, future UEs are
suggested where typical packet data applications should run without a standard
TCP at the radio interface. The plain UMTS radio interface protocol stack
offers reliable radio bearers with sufficient QoS for WWW browsing sessions.
To face the requirements of the Internet world, TCP might terminate or start
running end-to-end in the UMTS core network.
Biographies
Silke Heier and Andreas Kemper are Ph.D students at the Chair of Commu-
nication Networks. Together with her students, including Sebastian Grabner
and Jan-Oliver Rock, Mrs. Heier provided significant improvements to the
UMTS protocol simulator URIS. While she is almost finished with her degree,
mainly her students implemented layers two and three under her supervision.
Andreas Kemper is her successor with respect to further developments on the
simulator. Due to the current status of implementation, his focus is mainly
on the development of physical layer and radio resource control (RRC) rou-
tines. The horizon in development appears to be the coupling of this protocol
simulator with a system-level simulator to provide more detailed information
on channel quality, resulting for instance from propagation and interference
situation.
Acknowledgment
The authors would like to thank Prof. B. Walke of the Chair of Communi-
cation Networks for his support and friendly advice to this work.
250
References
25.322, G. T. (2001). Radio Link Control (RLC) Protocol Specification. Technical Specification
V 4.1.0, Release 4, 3rd Generation Partnerschip Project, Technical Specification Group Radio
Access Network.
Arlitt, M. F. and Williamson, C. L. (1995). A Synthetic Workload Model for Internet Mo-
saic Traffic. Proceedings of the 1995 Summer Computer Simulation Conference, Ottawa,
Canada, July, pp. 24-26.
Fall, K. and Floyd, S. (1996). Simulation-based Comparisons of Tahoe, Reno, and SACK TCP.
Lawrence Berkeley National Laboratory.
Frost, V. S. and Melamed, B. (1994). Traffic Modeling For Telecommunications Networks.
IEEE Communications Magazine. pp. 70-81.
Grabner, S. (2001). Simulative Performance Evaluation of the UMTS Radio Link Control Pro-
tocol. Master's thesis, Chair of Communication Networks, RWTH Aachen University of
Technology.
Holma, H. and Toskala, A., editors (200 I). WCDMA for UMTS: Radio Access for Third Gener-
ation Mobile Communications. John Wiley & Sons Ltd.
Mathis, M., Mahdavi, 1., Floyd, S., and Romanow, A. (1996). Selective Acknowledgement Op-
tions. RFC, Internet Engeneering Task Force.
Paxson, V. (1994). Empirically-Derived Analytic Models of Wide-Area TCP Connections. IEEEI
ACM Transactions on Networking, 2 (4), pp. 316-336, August 1994;
ftp:l/ftp.ee.lbl.govlpapersIWAN-TCP-models.ps.Z.
Rock, J.-O. (2001). Simulative Performance Evaluation ofIP based Services over UMTS. Mas-
ter's thesis, Chair of Communication Networks, RWTH Aachen University of Technology.
Steppler, M. and Lott, M. (1997). SPEET - SOL Performance Evaluation Tool. 8th SDL Forum
'97.
Walke, B. (2001). Mobile Radio Networks, 2nd Edition. John Wiley & Sons Ltd., Chichester,
England.
Interactions between the TCP and RLC
Protocols in UMTS
I. Introduction
Due to the rapid advances in the area of wireless communication and
Internet, provision of data services for applications such as e-mail, web
browsing, telnet, etc., over a wireless network is gaining importance. Because of
wireless channel constraints (mainly time varying conditions and bandwidth
limitation), the efficiency of the transmission protocols defined and optimized
for wired networks suffer when applied directly to wireless networks.
One important objective of UMTS (Universal Mobile Telecommunication
System) is to offer an Internet connection to a mobile subscriber by using an
efficient and reliable packet transmission protocol stack. The 3G- system will
provide for users higher data bit rates (around 384 kbps for mobile users and up
to 2 Mbps for indoor mobile users) than their 2G counterparts.
The reliable data transfer on the radio interface of UMTS, i.e. between an
user equipment (UE) and a Radio Network Control entity (RNC), is ensured by
the Radio Link Control (RLC) protocol. The RLC layer and the MAC layer
(Medium Access Control) manage QoS parameters required by the services.
In Internet, the dominant protocol used for end-to-end reliable data transfer
is the Transmission Control Protocol (TCP). The TCP protocol has been tuned
to perform well in wired networks where bit error rates are very low and
congestion is supposed to be the primary cause of packet losses. However, there
are some design issues in TCP, which make it difficult to use efficiently over the
251
X. Lagrance and B. fabbari (eds.).
Multiaccess, Mobility and Teletrcif.ficfor Wireless Communications, Volume 6,251-262.
© 2002 Kluwer Academic Publishers.
252
wireless links. In recent years, there have been large research activities in order
to improve the TCP performance in these networks (e.g., [1], [6]).
The suggested approaches can be classified in two categories ([4]). The
approaches in the first category try to hide non-congestion related losses from
the TCP sender (e.g., [3], [5]). They do not imply any changes in the existing
sender implementation. The intuition behind these approaches is that since the
problem is local, it should be solved locally. TCP does not have to be aware of
the characteristics of the individual links. In the corresponding solutions, most
of the losses, detected by the TCP sender, are caused by congestions. The
approaches in the second category attempt to make the sender aware of the
existence of wireless links and realize that some packet losses are not due to
congestion in the wired network (e.g., [6], [7]). The TCP sender can then avoid
invoking the congestion control algorithms when detecting non-congestion
related loss. A comprehensive comparison between different solutions for
improving TCP performance over wireless channels can be found for instance in
[4].
This paper intends to focus on the use of TCP over the UMTS radio
interface. The functionalities of TCP and RLC protocols are compared and
possible interactions between these two protocols are studied. A metric
indication concerning the RLC buffer occupancy is analyzed and we discuss the
impact of retransmissions at the RLC level on the TCP performance.
The paper is organized as follows. The next section contains a rapid
description of the radio protocol stack of UMTS. Section III develops a
communication scenario using TCP in the higher layer. The comparison
between the RLC and TCP protocols is the subject of section IV. Results of
simulation are discussed in section V. In the last section, our conclusions are
presented.
°
Access Control layer (MAC) via transport channels. The information of
transport channels are transmitted on radio frames of 1 ms. A small multiple of
the radio frame duration can be used to transmit data from radio transport
channel (i x 10 ms, where i E {O, 1,2, 3}).
The second layer (link layer) is split into two sub layers, RLC and MAC.
The MAC layer offers services to the RLC layer via logical channels. MAC
layer handles different types of logical channels with different QoS parameters
and supports fast adaptation mechanisms provided by the Radio Resource
Control layer (RRC). RLC entities handle Protocol Data Units (RLC PDUs or
blocks). RLC blocks can transport user data (e.g. a TCP segments) or it can
transport signaling messages (for instance RRC messages).
253
'I
FIl' FIl'
IITIP IfITP
In wireless networks the packets losses are produced for other reason than
congestion. The losses are mainly due to high bit error rates (as high as 10·3) and
handoff procedures. These lost packets are interpreted by TCP as network
congestion event and therefore TCP invokes for each event the congestion
mechanisms. This misinterpretation results in an unnecessary reduction in end-
to-end throughput and the TCP performance degradation.
In cellular systems, there are defmed protocols that implement mechanisms
that allow rapid detection of erroneous blocks transmitted over the radio
interface. These protocols use mechanisms such as ARQ, FEC (Forward Error
Correction) or hybrid ARQIFEC.
called STATUS PDD. The piggybacking mechanism can be used. In this case, a
special control block named Piggybacked STATUS PDU is defmed.
Both TCP and RLC can change dynamically the transmission window
during a communication. For TCP, the current transmitter's window size txwnd
is the minimum between the congestion window size cwnd and the advertised
reception window size rxwnd. In case of RLC, the receiver can change the
transmission window by sending STATUS PDU with a new size of the window
(by using the "Super Field Window" described in [13]).
A RLC entity can transfer data by using either one logical channel or two
logical channels. If it uses one logical channel, the size of a RLC control block
(STATUS PDU) must have the same size as the RLC block for data (PDU).
The unused space in the control block is filled by padding. In order to optimize
the size of the blocks for data and control, RLC makes it possible to use two
logical channels. One channel is used for transport of data and the second
channel for transport of control.
Due to data services that TCP support (Web browsing, file transfer etc.),
the downstream and upstream traffics are asymmetric. The size of a TCP
segment containing data is much higher than the size of a TCP segment with
Ack. This feature has an impact on the radio interface. In order to use the radio
resource efficiently, it is desirable to allocate channels with different capacity
for the downlink and the uplink connection. This data rate ratio has do be
determined carefully. Otherwise, the adverse effects on the TCP performance
can take place when this ratio is not appropriate (e.g., [2], [10], [12]). Table 1.
sunnnarizes some of the features ofTCP and RLC.
RLC TCP
Concerned link MS-RNC End-to-end
Go-Back-N
fast retransmission,
ARQ mechanisms Selective Repeat
fast recovery
SACK (optional)
SNmodulo PDUs (blocks) Bytes
Numbering AM: i _1
2
232 - I
(Sequence Number) UM: 27-1
Tr: unidirectional
Mode UM: unidirectional Bi-directional
AM: bi-directional
Piggybacking Piggybacking
How to send an Ack
STATUSPDU Frame with an Ack
Tr: '" 0 to 80
Aver. size of unit
UM:",20 '" 500 to 1500
[Bytes]
AM:",16t080
V. Simulation
The simulation is carried out on a RLC/TCP test-bed developed at ENST.
The architecture of the test-bed is described in figure 2. The test-bed comprises
2 PC that are connected via Ethernet. The first PC is used as UE and the second
PC simulates RNC entity, wired network and Server. The data transmission
between these PCs is based on the socket interface via Ethernet using a
tunneling protocol. The time of simulation is controlled by Supervisor in the
UE.
In our analysis we consider an uplink data transmission (file transfer). The
main parameters used during the simulations are given in table 2. The average
round trip time in the wired network is equal to 200 ms and the standard
deviation is equal to 20 ms ([9]). The round trip time in the wireless network,
i.e. between UE and RNC entity, is set to 20 ms when no retransmission occurs.
--------------------------------------., ,,------------------------------------------,,
: : Superviso '
VE
: : : j
,______________________________________ 2 ,.------------------------ ----------------,
medium
where C denotes the data rate of the radio channel and RTTwireless (respectively
RTTwired) is the mean value of round trip time in the wireless network
(respectively in the wired network).
At the end of the phase CgAvl, the TCP parameter txwnd reaches the
maximum value (in our example, it is set to 40 MSS); then it stays constant.
Likewise, the RLC buffer occupancy stops increasing linearly and oscillates
258
around a mean value (phase CgAv2 in figure 3a). In the graph, its value is
around 23 SDUs.
A congestion event, introduced by a segment loss in the wired network (Er
in figure 3a), can be observed at time 7s. TCP detects the congestion event by
receiving three duplicated Acks. When receiving the third duplicated Ack, TCP
retransmits the segment loss and reduce its congestion window by one half.
Subsequently, the TCP sender enters into the fast recovery phase. During this
phase, the TCP sender "inflates" its window by the number of DupAcks
(duplicated Acks) it receives. Each received DupAck indicates that a segment
has been removed from the network and is now cached at the receiver side. The
sender waits until a certain number of DupAck are received without delivering
to the RLC entity any segments. During this period, the buffer occupancy
dramatically decreases because (i) there are no new delivered SDUs from TCP
and (ii) the RLC entity continues to serve queued SDUs in the buffer.
Once the TCP sender has received a fresh Ack, it exits the fast recovery
phase and enters into the congestion avoidance phase. The buffer occupancy
starts again to growth linearly (phase CgAvl' in figure 3a).
Figure (3b) shows the buffer occupancy for a radio channel with
retransmission of RLC blocks. Due to the retransmissions, the buffer occupancy
can reach during certain periods a much higher value than in figure (3a). The
RLC buffer size should be chosen with respect to rxwnd (TCP receiver window
size) and to the BIER on the radio interface.
10 15
Timels1
Figure 3a. Number of SDUs in the RLC buffer as function of time: error free
radio channel. There is one TCP segment loss introduced in the wired
network (at time around 7s).
1Q 15
Timels)
Figure 3b. Number of SDUs in the RLC buffer as function of time: radio channel
with errors (BIER=10%). As in figure (3a), a TCP segment loss is
introduced.
260
Figure 4a. Transmission delay over the radio interface for each SDU: error free
radio channel. As in figure (3a), a TCP segment loss is introduced.
Figure 4b. Transmission delay over the radio interface for each SDU: radio
channel with errors (BIER=10%). As in figure (3a), a TCP segment
loss is introduced.
261
VI. Conclusion
This paper has investigated some issues that emerge when using TCP over
the RLC protocol in UMTS. We have compared and analyzed the protocol
mechanisms used by TCP and RLC in acknowledged mode. Some mechanisms
can be found in both protocols: segmentation/reassembly, flow control and
ARQ. We have noticed that the RLC mechanisms are more adaptable to the
radio constraints than those used in TCP.
In our simulation, we have considered a single uplink TCP connection. We
have shown that the RLC buffer occupancy reflects the different TCP phases:
slow start, congestion avoidance, fast retransmit and fast recovery. The
dimensioning of the RLC buffer should be done with regard to the TCP receiver
window size and to the radio interface BIER.
This study has shown reciprocal interactions between the protocols. We
presently investigate some possibilities of coupling TCP and RLC mechanisms.
References
[1] E. Ayanoglu, S. Paul, T. F. LaPorta, K. K. Sabnami, and R. D Gitlin,
..AIRMAIL: A link-layer protocol for wireless networks," ACM
ACMIBaltazer Wireless Networks Journal, February 1995.
[2] Y. Bai, A. T. Ogielski, "TCP over asymmetric CDMA radio links," Proc.
IEEE Con! on Vehicular Technology (VTC 2000 Fall), September 2000.
[3] A. Bakre, B.R. Badrinath, "I-TCP:Indirect TCP for mobile hosts," Int.
Con! Distributed Computing Syst. (ICDCS), May 1995.
[4] H. Balakrishnan, V. Padmanabhan, S. Seshan, and R. H. Katz, "A
comparison of mechanisms for improving TCP performance over
wireless links," ACM SIGCOM'96, August 1996.
[5] H. Balakrishnan, S. Seshan, and R. H. Katz, "Improving reliable
transport and handoff performance in cellular wireless networks," ACM
Wireless Networks, December 1995.
[6] S. Biaz, N. H. Vaidya, "Distinguishing congestion losses from wireless
transmission losses: a negative result," Computer Communications and
Networks, 1998.
[7] W. Ding and A. Jamalipour, "A new explicit loss notification with
acknowledgment for wireless TCP," Personal, Indoor and Mobile Radio
Communications, September 2001.
[8] H. Holma, A. Toskala, "WCDMA for UMTS, Radio Access for Third
Generation Mobile Communications," John Wiley & Sons, Ltd, England
2000.
262
[9] T. J. Kostats, M. S. Borella, I. Sidhu, G. M. Schuster, J. Grabiec, J.
Mahler, "Real-time voice over packet-switched networks," IEEE
Network, Jan.-Feb. 1998.
[10] T. V. Lakshrnan, U. Madlow, B. Suter, "Window based error recovery
and flow control with a slow acknowledgment channel: a study of
TCPIIP performance," Proc. INFOCOM, April 1997.
[11] R. Ludwig, R. H. Katz, "The Eifel algorithm: Making TCP robust against
spurious retransmission," ACM Computer Communication Review, Vol.
30, No.1, January 2000.
[12] S. Varna, "Performance and Buffering Requirements ofTCP Application
in Asymmetric Networks," Proc. INFOCOM, March 1999.
[13] 3G TS 25.322 (2001-06), RLC protocol specification (Release 4).
IMPACT OF SR-ARQ WITH FINITE
BUFFER ON TDDjTDMA WIRELESS LAN
Emilio Strinati,
Jeremy Gosteau,
Sebastien Simoens
and Pietro Pellati
Motorola Labs Paris, Saint Aubin 91193 Gi/-sur- Yvette France
tel. :+33-{0)1-69-35-25-64
fax. :+33-(0) 1-69-35-25-01
jeremy.gosteau@crm.mot.com
Abstract In this paper, the influence of some implementation parameters of Selective Re-
peat Automatic Repeat Request (SR-ARQ) on system performance is investi-
gated. In the framework of the specific SR-ARQ algorithm specified by the
ETSI BRAN HIPERLAN12 (H/2) Wireless LAN standard, the need for optimiz-
ing the ARQ signalling bandwidth is illustrated and several signalling strategies
are presented. Even with optimum management of the signalling bandwidth, the
finite transmit and receive buffers can seriously limit the throughput. This effect
is modeled by using a simple probabilistic approach, relying on the TDDITDMA
access scheme, and is evaluated by simulation. The interaction of SR-ARQ with
scheduling and Link Adaptation is also discussed and finally, an ARQ aware
scheduling strategy is proposed.
1. Introduction
Any communication system implements mechanisms for limiting the trans-
mission of erroneous messages. These techniques can be sorted in two main
categories depending on their using error-correcting or error-detecting codes.
In the latter, if a receiver detects an erroneous packet, it sends a message back
to the transmitter to signal the error.
The scheme used by such a system is called Automatic Repeat Request
(ARQ). Many instanciations have already been thoroughly studied (Lin and
Costello, 1983), (Gibson, 1997). They range from the stop-and-wait to the se-
263
X. Lagrance and B. fabbari (eds.),
Multiaccess, Mobility and Teletraffic for Wireless Communications, Volume 6, 263-278.
© 2002 Kluwer Academic Publishers.
264
lective repeat (SR) algorithms. This paper just focuses on the latter. Indeed,
the SR scheme provides efficient retransmission with a limited overhead of ac-
knowledgment and a reasonable delay: only the erroneous packets which are
negatively acknowledged or for which the time out has expired are repeated.
This implies that buffers must be provided at both the transmitter and receiver
side in order to store the not yet positively acknowledged packets.
In this paper, three issues related to the SR-ARQ are addressed: the sig-
nalling strategy, the limitations due to the finite transmitting or receiving buffer
and the impact of the scheme on the scheduling and on the Link Adaptation
(LA).
In (Li et al., 2000), the authors study the signalling mechanism of an SR-
ARQ scheme in the framework of Hl2 Wireless LAN. They show that, in the
case of a downlink (DL) connection, an incremental allocation of the ARQ ac-
knowledgment messages optimizes the throughput on the Data Link Control
(DLC) layer. In the case of an uplink (UL) connection, we show in section
2 that the algorithms can be further improved if the exact required number of
feedback messages is dynamically granted. We then compare various algo-
rithms for DL or UL connections, and select the most efficient one. With this
first step, a strategy is proposed to optimize the throughput with respect to the
signalling constraints in the case of an Hl2 based network.
Nevertheless, a limitation in the throughput can still be observed even for
large buffer sizes when the Packet Error Rate (PER) grows. In section 3, we
model the phenomenon with simple discrete probabilities calculus, relying on
a TDDffDMA access scheme, which differs by the approach and the assump-
tions from what can be found in (Miller and Lin, 1981), (Saeki and Rubin,
1982) or (Jianhua et al., 1999). The analytical approach is compared with re-
sults simulated with a H12 network simulator.
To complete the study of the SR-ARQ scheme with a finite buffer size, sec-
tion 4 evaluates its impact on the scheduling and the LA. Actually, the schedul-
ing of resources has already been studied in papers like (Kadelka et al., 1999)
or (Ranasinghe et al., 2001) for TDDffDMA based systems. These studies are
yet limited to alternatives of round robin algorithms without involving ARQ
parameters. We propose here to explain to what extent the choice of some
ARQ parameters can greatly influence the choice of a scheduling algorithm and
thereof the resulting overall system performance. We will show that a metic-
ulous setting of the parameters is key to avoid a drop in throughput. Simple
guidelines can be drawn out of this study. Based on these results, a scheduling
strategy is proposed. In the same way, the Link Adaptation (LA) (Goldsmith
and Chua, 1998), (Simoens and Bartolome, 2001) needs some slight tuning in
order to take the ARQ into account. This issue is also discussed in section 4.
Impact of SR-ARQ with finite buffer on TDD/TDMA Wireless LAN 265
.. ..
Broadcast
POUTrain
- OL
POUTrain
..
DL
POUTrain
..
UL
POUTrain
- --
UL UL
POU Train POU Train
The PDU trains which are of concern here consist of SCHs (see table I)
to carry control information and LCHs mainly used to carry payload. In the
direct link phase (that will not be considered any more in what follows), Mobile
Terminals (MTs) send PDU trains to each others in a peer-to-peer manner.
During the DL phase, the access point sends PDU trains either in multicast or
266
to a specific MT. Lastly, MTs send PDU trains to their access point during the
UL phase. Figure 1 illustrates this.
A connection is identified with a DLC User Connection IDentifier (DUC
ID) and a given MT can manage several connections either DL or UL:
- if a DL connection is considered, the access point sends LCHs contain-
ing payload to the MT and the latter acknowledges the receipt of the
packets in SCHs containing ARQ feedback messages. In these mes-
sages, the MT can request more bandwidth for acknowledgement using
the ABIR bit (ARQ Bandwidth Increase Request);
- if an UL connection is considered, the MT sends the payload in LCHs
and the acknowledgement is done in SCHs sent by the access point.
In both cases, as the resource allocation is centralized, the scheduler in the
access point grants the number of LCHs and SCHs for each connection. To
get the resources, the MT makes its request via an RCH or an SCH. Thus, for
limiting the overhead and for the buffer management, an adequate choice of
the signalling strategy is key.
This section deals with the choice of the strategy. This will be Hl2 ori-
ented but the algorithms can be extended to any other system relying on a
TDDrrDMA access scheme.
The ABIR based algorithm is close to the one proposed in (Li et aI., 2000).
Indeed, when the access point receives an ARQ feedback message with the
ABIR bit set, the number of SCHs is increased by one the frame after with no
upper limit. Otherwise, this number is decreased by one. Yet, this value is kept
greater than one.
In the case of the second algorithm, the scheduler grants all the needed ARQ
feedback messages to the connection up to a given limit denoted Max. We
Impact of SR-ARQ with finite buffer on TDD/TDMA Wireless LAN 267
tested this scheme for Max ranging from 1 to 3. This limits the overhead with
a highly dynamic bandwidth.
The third algorithm is an extension of the second with no upper limit. This
may grow the overhead but the scheme responds to any variation of the traffic
with no delay.
The performance of these three algorithms is compared in terms of through-
put calculated on top of the DLC layer (i.e. provided to the IP layer). Before
dealing with the simulation results, let first derive an expression for the ideal
throughput.
'toverhead)
P[Mbps] = r mode[Mbps]' ( I - A
. 1-"
(
I - PER
)
(1)
'tframe
where
- rmode is the nominal bit rate ranging from 6 to 54 (Mbps) depending
on the physical mode selected for the transmission of the LCHs (see in
(ETSIIBRANIDLC, 2000))
- 'toverhead is the part of the MAC frame containing no payload. Referring
to (Kadelka et al" 1999) and (ETSIIBRANIDLC, 2000), this is worth
The duration of the SCHs also depends on the physical mode used by the
LCHs. Also note that the MAC overhead includes propagation delays
(guard times).
- 't frame equals 2 ms
- ~ represents the overhead ratio introduced by the Cyclic Redundancy
Check (CRC) bits (ETSIIBRANIDLC, 2000) and the convergence layer
header (ETSIIBRAN/CL, 2000). This value is worth
~= 48 (3)
54
- PER is the Packet Error Rate ranging from 0 to 20% in our simulations,
which corresponds to typical operating conditions.
Equation (1) reflects ideal ARQ assumptions that is an infinite buffer size,
an unlimited number of retransmissions, error-free signalling and no signalling
bandwidth limitation. Ideal ARQ assumptions are not realistic but will provide
268
an upper bound to the throughput. Finally, packet errors are assumed inde-
pendent. In the H/2 context, such an assumption is valid in noise-limited en-
vironments, where thermal noise produces bit error bursts at the output of the
Viterbi decoder which are much shorter than the packet length.
5 10 IS ro ~ ~ ~ ~ ~ M ~ ~ ~ ~ n ~ ~ ~ ~ 100
a..MIW{%)
~ :.'!
I""
l
k=J
C2 .
Cl
Figure 4a. Non Exhaustive Round Figure 4b. Exhaustive Round Robin
Robin Algorithm (NERR) Algorithm (ERR)
- best-SNIR ERR: the connection having the best Signal to Noise plus
Interference Ratio is served first - this implies a high throughput but the
slowest connections may never be served;
- time-based NERR: the scheduler allocates the same duration for each
connection no matter what their modulation is - even the slow connec-
tions will be served;
- data-based NERR: the scheduler allocates the same amount of data for
each connection - this will provide the fairest algorithm at the expense of
the cell throughput. In the following NERR will stand for this algorithm.
Fairness Cell throughput
..
t t t
data-based time-based best-SNIR
NERR NERR ERR
Let now see how we can choose a scheduling algorithm based on the ARQ
configuration used.
versus PER with a fixed ARQ window size set to 512 and served by NERR. The
total throughput almost reaches the ideal ARQ upper-bound. As a reference,
the throughput obtained in the same conditions but with a single connection is
also plotted. The latter can be viewed as a "worst case" of what can be reached
in the multi-connection case when only one connection is served per frame,
and all connections are stalled simultaneously. Basically, NERR performs very
well because the ratio a (cf. section 3) was divided by two.
Now this phenomenon has been clearly highlighted, let see how this trans-
lates into recommendations for tuning scheduling algorithms. For that purpose,
let consider simulation results plotted on figure 3. If a throughput efficiency
greater than 98% is imposed, with a PER of 10%, a needs to be less than
28%. For simplification sake, the scenario is restricted to n identical connec-
tions, each set to the same physical mode <p (from 0 for BPSK rate ~ to 6 for
64 - QAM rate i). If each frame is completely filled, the number of LCHs per
frame is thus related to the physical mode. A relation between <p, n and the
window size W can therefore be derived as represented on figure 7.
This graph can be read in the following manners:
- if we have one user in the cell, we cannot have a W smaller than or equal
to 64;
274
512
X maximum PHY mode (O=BPSKII2 -> 6=64QAM3/4)
.~
nb users
Figure 7. Maximum physical mode to use for a given number of users and a given window
size
- if we have several users in the cell, all in physical mode 5 (36 Mbps) and
with a W of 128, then this number of users must be greater than or equal
to 5;
- if all the users have a W of 32 and a physical mode 6 (54 Mbps), there
needs to be at least 26 users in the cell;
- if 4 users share the cell with the 6th physical mode (54 Mbps), W has to
be greater than or equal to 256;
- if 7 users share a cell with a W of 64, then their physical mode should
not exceed 4 (27 Mbps);
- if all connections have a physical mode 5 and a W of 64, the optimum
number of users in the cell using a NERR algorithm is 9.
These sub-groups being served by a fair ERR, in which the first served group
changes cyclically. For clarification sake, let consider the simplistic hypothesis
which led to figure 7, if all users have a window size of 128 and are transmitted
in physical mode 5, Aequals 5 and the optimum number of users in the cell can
be taken equal to 5. Note that such an algorithm is not simulated in this paper
and is currently under evaluation.
~J~
ERR
5. Conclusion
In this paper, two SR-ARQ signalling strategies well adapted to TDD/-
TDMA access based systems either for an UL or for a DL connection are
proposed. Nevertheless, a gap is observed between the theoretical throughput
and the one obtained by simulation in a Hl2 network. We then develop an an-
alytical approach to derive a new formula for the throughput which takes the
finite buffer into account. By doing so, we verify that finite buffer space is a
major factor limiting the throughput of the SR-ARQ scheme. We also show
that the throughput loss can be recovered by carefully setting ARQ buffer size
of each connection, modifying the link adaptation switching points and care-
fully designing resource allocation algorithm.
Impact of SR-ARQ with finite buffer on TDD/TDMA Wireless LAN 277
No
p(A) = L p(Aj} (A.l)
j=1
with A j the event: « The mth packet of the block is not received correctly until the /h attempt
(which is successful) ». We have p(Aj} = Ej-l(1- E) where E denotes the PER. Thus (A.I) can
be rewritten as:
No. I_ENo
p(A) = (I-E) L EJ - 1 = (I-E)-- = I_ENo (A.2)
j=l I-E
The event (N = No and n = no) can be expressed as BnCnD with:
- B: «the first no - 1 packets of the block are correctly received at the latest at the (No -
l)fh transmission»
p(B) = (1- ~o-l )no-l
- C: «the n~ packet is correctly received at No and not before »
p(C) = E 0-1(I- E)
- D: «the M - no remaining packets are correctly received at the latest at No »
p(D) = (I_ENO)M-nO
Since packet errors are independent, events B,C and D are also independent. Therefore, PNo,no =
p(B)p( C) p(D). Having PNo,no' the expression of X is direct:
The equation A.3 reflects the three conditions necessary to produce a transmission stall as de-
scribed in section 3. y(No) is the probability of the second event: «There is no block in the buffer
older than No ». This means that the block sent just before the considered block was received
correctly at age No and that the block sent before the previous one was received correctly at age
No + I and so on. Since blocks are independent just like packet errors arc, y(No) is the product
of these probabilities as written in A.4.
278
References
ETSIIBRAN/CL (2000). Broadband Radio Access Networks (BRAN); HIPERLAN Type 2
Functional Specification; Packet based Convergence Layer; Part 1: Common Part. Norme
ETSI, document TS101493-1 vl.l.l, European Telecommunications Standards Institute,
Sophia-Antipolis, Valbonne, France.
ETSIIBRAN/DLC (2000). Broadband Radio Access Networks (BRAN); HlPERLAN Type 2
Functional Specification; Data Link Control (DLC) layer; Part I: Basic Data Transport Func-
tion. Norme ETSI, document TSI01761-1 v1.2.1, European Telecommunications Standards
Institute, Sophia-Antipolis, Valbonne, France.
Gibson, J. D. (1997). The Communications Handbook. CRC Press, IEEE Press.
Goldsmith, A. J. and Chua, S. (1998). Adaptive coded modulation for fading channels. IEEE
Trans. on Communications, 46:595-602.
Jianhua, H., Zongkai, Y., and Donghua, D. (1999). Performance analysis on selective repeat
ARQ with finite buffer. IEEE Trans. on Communications, 2:1101-1104.
Kadelka, A., Hettich, A., and Dick, S. (1999). n Performance Evaluation of the MAC protocol
of the ETSI BRAN HIPERLAN/2 standard n. Proceedings of the European Wireless '99.
Li, H., Lindskog, J., G.Malmgren, Miklos, G., Nilson, F., and Rydnell, G. (2000). n Automatic
Repeat Request (ARQ) Mechanism in HIPERLAN/2 n. Proceedings of the IEEE VTC Spring
conference.
Lin, S. and Costello, D. J. (1983). Error Control Coding: Fundamentals and Applications. En-
glewood Cliffs, NJ: Prentice-Hall.
Lin, Z., Malmgren, G., and Torsner, J. (2000). System Performance Analysis of Link Adaptation
in HIPERLAN Type 2. In Proceedings of the IEEE VTC Fall conference, volume 4, pages
1719-1725.
Miller, M. J. and Lin, S. (1981). The analysis of some selective-repeat ARQ schemes with finite
receiver buffer. IEEE Trans. on Communications, 29: 1307-1315.
Ranasinghe, R. S., Andrew, L. L. H., Hayes, D. A., and Everitt, D. (2001). Scheduling disci-
plines for multimedia WLANs: Embedded round robin and wireless dual queue. In Proceed-
ings of the IEEE ICC Conference, volume 4, pages 1243-1248.
Rosberg, Z. and Sidi, M. (1990). Selective Repeat ARQ: the joint distribution of the trans-
mitter and the receiver resequencing buffer occupancies. IEEE Trans. on Communications,
38: 1430-1438.
Saeki, B. H. and Rubin, I. (1982). An analysis of a TDMA channel using stop-and-wait, block,
and select-and-repeat ARQ error control. IEEE Trans. on Communications, 30: 1162-1173.
Simoens, S. and Bartolome, D. (2001). Optimum performance of link adaptation in HIPER-
LAN/2 networks. In Proceedings of the IEEE Vehicular Technology Conference, Rhodes,
Greece.
Traffic Performance Analysis of Multimedia
Applications in Evolved GSM Networks
Abstract This paper presents traffic performance results for different traffic mixes in cellu-
lar packet radio networks. These results are useful to estimate the radio capacity
that is needed during the evolution of GSM networks towards third-generation
(3G) mobile communication systems like GSMIEDGE Radio Access Networks
(GERAN). First a traffic mix of applications based on the Wireless Application
Protocol (WAP) and conventional Internet applications like WWW browsing and
e-mail over the General Packet Radio Service (GPRS) are regarded. In the next
step traffic performance results for Streaming applications over GPRS and EG-
PRS are presented and the feasibility of Streaming with coexisting interactive
and background applications like WWW and e-mail is examined. Simulation
results for quality of service measures for the different applications and GPRS
system measures are based on the simulation tool GPRSim that models the appli-
cation and user behavior, the TCPIIP and WAP protocol architecture, the GPRS
protocol architecture and the radio channel.
1. Introduction
The driving force for the evolution of second generation mobile communi-
cation systems such as the Global System for Mobile Communication (GSM)
is the predicted user demand for mobile data services that will offer mobile
Multimedia applications and mobile Internet access.
After High Speed Circuit Switched Data (HSCSD) has been introduced in
some countries in 1999, the first GPRS-based services have been available
since 2001 in Europe. Many countries worldwide will introduce GPRS in the
next years. With these new services mobile data applications with net bit rates
of up to 117 kbitls will be offered and established on the market. To real-
ize higher data rates the European Standardization Institute (ETSI) and the
3rd Generation Partnership Project (3GPP) have developed the Enhanced Data
279
x. Lagrance and B. Jabbari (eds.),
Multiaccess, Mobility and Teletraffic for Wireless Communications, Volume 6, 279-294.
© 2002 Kluwer Academic Publishers.
280
Rates for GSM Evolution (EDGE) standard, which offers a net bit rate of up
to 384 kbitls by means of modified modulation, coding and medium access
schemes (see Furuskar et aI., 1999; Stuckmann and Franke, 2001).
In parallel to the GSM evolution, the data applications performed by mo-
bile users will evolve. In the first phase of the GSM evolution, where the data
services Circuit Switched Data (CSD) and GPRS are available, WAP-based
applications as defined in Wireless Application Protocol Forum, 1999 running
on smart phones and PDAs besides conventional Internet applications running
on laptop computers or enhanced PDAs will dominate. Then Video Stream-
ing applications (see Elsen et aI., 2001) and Large Data Transfer (LDT) ap-
plications including the Multimedia Message Service (MMS) based on WAP
version 2.0 as defined in Wireless Application Protocol Forum, 2001 are felt to
become more popular with the optimization of GPRS and with the introduction
of EDGE and the related packet data service Enhanced GPRS (EGPRS).
While for the time period right after the service introduction minimal con-
figurations were chosen supporting only a basic availability of GPRS, with
increasing data traffic load in the next years GSMlGPRS cell capacity will
have to be extended. For this evolution of GSMlGPRS networks and for
the introduction of EGPRS, dimensioning guidelines are needed for operators,
equipment manufactures and system integrators. They should describe the re-
lationship between the offered traffic and the radio resources to be allocated
to reach a desired quality of service for the different applications (see Walke,
2001; Stuckmann and Paul, 2001; Stuckmann, 2002).
This paper aims at presenting simulation results for two predicted traffic
mixes, one for a GPRS evolution scenario and one for an EDGE introduction
scenario. The first one is composed of WAP, WWW and e-mail, the second is
defined by Streaming, WWW and e-mail.
In Section 2 the potential applications and the related traffic models are
introduced. After the description of the simulation tool GPRSim in Section 3
the traffic performance results are presented and interpreted in Section 4.
www
All applications summarized by World Wide Web (WWW) are based on the
Hypertext Transfer Protocol (HTTP), which uses the TCPIIP protocol stack.
HTTP organizes the transfer of Hypertext Markup Language (HTML) docu-
ments (web pages).
281
E-mail
E-mails are transmitted by using the Simple Mail Transfer Protocol (SMTP)
or the Post Office Protocol version 3 (POP3) for e-mail download. Since the
size of an e-mail download on a mobile device is the crucial parameter for this
research, a traffic model defining e-mail sizes is suitable. The introduced e-
mail model based on Paxson, 1994 describes the load arising with the transfer
of messages performed by an SMTP user. The only parameter is the e-mail size
that is characterized by two log2-normal distributions plus an additional fixed
quota of 300 byte (see Table 1). The base quota was assumed to be a fixed
overhead. Subtracting the overhead, a bimodal distribution remained. The
lower 80 % were said to be text-based mails, while the upper 20 % represent
mails with attached files, which can be rather large. The transition between
these two distributions is 2 kbyte. The maximum e-mail size is set to 100 kbyte.
WAP
The WAP specifications, which are the basis for the implementation in to-
day's mobile terminals, including the June 2000 Conformance Release, also
known as WAP 1.2.1, aim at optimizing the operation in 2G networks. There-
fore WAP 1.2.1 defines a distinct technology comprising protocols and content
representation. WAP is a suite of specifications that defines an architecture
framework containing optimized protocols (e.g., WDP, WTP, WSP), a com-
pact XML-based content representation (WML, WBXML) and other mobile-
specific features like Wireless Telephony Applications (WTA) as defined in
Wireless Application Protocol Forum, 1999.
protocol stacks will still be used in the mobile terminals in the next years. In
this paper, only WAP l.x is regarded.
WAP Traffic Model. A WAP traffic model has been developed and
applied in Stuckmann et aI., 2001; Stuckmann and Hoymann, 2002.
A WAP session consists of several requests for a deck performed by the
user. The maximum amount of data that can be transferred by one request
defaults to 1400 bytes. The parameters are summarized in Table 1. The main
characteristic is a very small mean packet size (511 byte) modelled by a log2-
normal distribution with a limited maximum packet size of 1400 byte (see
Table 1).
Video Streaming
Many Internet portal sites are offering video services for accessing news and
entertainment content from a Personal Computer (PC). Beside Motion Picture
Expert Group (MPEG), H.263 is the currently most accepted video coding
standard for Video Streaming applications. In the near future, mobile commu-
nication systems are expected to extend the scope oftoday's Internet Streaming
solutions by introducing standardized Streaming services as described in Elsen
et aI., 2001.
In the scope of modelling video sources, a lot of attention has been paid to
long range dependent or self-similar models of traffic streams in telecommuni-
cation networks (see Willinger et aI., 1997). Many of such models have been
used to investigate Variable Bit Rate (VBR) video sources with a statistical
analysis of empirical sequences and estimation of the grade of self-similarity
(see Rose and Frater, 1994). Since MPEG and H.263 video traffic consists of
a highly correlated sequence of images due to its encoding, the correct mod-
elling of the correlation structure of the video streams is essential (see Zaddach
and Heidtmann, 2001).
In this work no stochastical models of video streams with self-similar or
high-correlated traffic characteristics are applied. Real video sequences coded
by an H.263 coder are used to generate the Streaming traffic.
The Video Streaming traffic model used within the scope of this work is
based on three video sequences in the format QCIF (Quarter Common Inter-
mediate Format) with the resolution of 176 x 144 pixels. The sequences are
proposed by the Video Quality Expert Group (VQEG) and are for this reason
commonly used. Each sequence is representing a particular group of videos
with different intensities of motion.
• Claire stands for a very low motion intensity and can be seen as a
characteristic video conferencing sequence or inactive visual telephony.
• Carphone includes both, periods with rather high motion and peri-
ods of low motion intensity. It represents many kinds of vivid or active
video-conferences or even visual telephony.
284
Beside visual telephony all of the new emerging applications are relatively
short in duration. So called heavy users, generating long streams with huge
amounts of data, have not been taken into account. The duration of video
sessions is modelled by a negative-exponential distribution with an average
value of 60 S. This is an assumption with regards to the prognosis for 3G
networks in ETSI 3GPP, 1998 where the duration of real-time calls is proposed
to be modelled by a negative-exponential distribution.
3. Simulation Environment
The full details of the GPRS protocol stacks of the radio and the fixed net-
work and of the Internet protocols including the characteristics of TCP cur-
rently cannot be described by formulas usable in practice. Since GPRS net-
works are presently introduced in the field, traffic engineering and related per-
formance results are needed soon, so that capacity and performance estima-
tions become possible for GPRSIEDGE introduction and evolution scenarios.
Measuring the traffic performance in an existing GPRS network is not pos-
sible, since a scenario with a well-defined traffic load is hard to set-up, the
285
IRallwayl IP
Generator
GIST
Mean application response time is the difference between the time when a
user is requesting a web page, a WAP deck or an e-mail and the time
when it is completely received.
Mean throughput per cell is also called system throughput and is calculated
from the total IP data transmitted on all PDCHs of the regarded radio
cell and for all users during the whole simulation duration, divided by the
simulation duration. Since a loss of IP datagrams over fixed subnetworks
is not modelled, this parameter equals the offered IP traffic in the radio
cell.
PDCH utilization: is the number of MAC blocks utilized for MAC data and
control blocks normalized to the sum of data, control and idle blocks.
Thus existing capacity reserves in the scenario under consideration can
be seen from this measure.
-----------------
•• .
:=::::==:::::~.::::::::::::::::=:=;======~==~=~~----------------
,
NumbefofMS
15 20
WWW4PDCH-
... e-mail 4 PDCH -------
............., WAP, 4 POOH --.... -.
I,.
t
;
0;. I,.
t -----~-~----
e:
..
~ 10 10
~ ~
NumberofMS
____ ~ ____ ~
~
. L -____- L____
•
~~
NurnbarofMS
____ ~ ____ ~
20
with 10 active MS in the radio cell even if 4 PDCHs are available. The reason
for the strong increase in response time for WWW and e-mail can be seen in
other evaluated measures like the downlink PDCH utilization in Figure 4(a).
100 % PDCH utilization is reached for WWW/e-mail traffic with 15 MS, while
15 WAP users are only utilizing the PDCHs with 30 % for the same PDCH
configuration.
289
Figure 3(a) shows the mean downlink IP throughput per user during trans-
mission periods. While the throughput performance for pure WAP traffic re-
mains relatively constant with an increasing number of mobile stations and
4 PDCHs, it decreases dramatically for pure WWW/e-mail traffic because of
the higher offered traffic and the higher utilization. The poor throughput per-
formance for WAP traffic can be explained by the low WAP deck size. Such
transaction-oriented applications are more influenced by the high round-trip-
time, which is mainly caused by the high delay over the air interface, than by
the available bit rate. Since the response time for a WAP deck is less than 1.5 s,
which should be acceptable for a wireless application, the user is not aware of
this low throughput performance.
Since WWW and e-mail applications comprise larger file sizes to down-
load than WAP-based applications do, the throughput performance perceived
by a user in situations with low traffic load ranges from 14 to 24 kbitls. These
performance values are mainly influenced by the characteristics of the offered
traffic. Since the e-mail traffic model has larger file sizes than WWW, the
throughput performance is better. With an increasing number of mobile sta-
tions up to 15 the saturation is reached and the performance for WWW and
e-mail users gets unacceptable and even gets worse than the low throughput
for pure WAP traffic. In this situation with high traffic load the WWW and
e-mail traffic performance is less influenced by the characteristics of the traffic
model like the file size, but by the load on the air interface.
80
I6 60
]
s
Ii
~ 40
I 20
°0L---~--~--~--~20
Number of MS Number of MS
(a) Pure WWW/e-mail and WAP traffic (b) Traffic Mix compared to pure WWW and
pure WAP traffic
8000
1
~ 6O<lO
-8
{•
E
4000
"-
5
10
2000
NumberofMS NumberofMS
(a) Mean downlink IP datagram delay (b) Mean downlink IP troughput per user
Probability distribution function (1 MS par call, video streaming) Probability disbbution function (10 MS pat eell, video streaming)
D.•
0.'
~
it:
0.'
0.2
0
0
DL IP datagram dalay [msJ
Figure 6. Distribution of the downlink IP datagram delay for Video Streaming applications
(traffic mix)
cell 15-20 users generating the traffic with 10 % Streaming can be satisfied.
The delay starts increasing dramatically with 15 and 20 users, respectively.
Regarding the downlink IP throughput per user in Figure 5(b) there is no
significant difference in the performance between 6 and 8 available PDCHs.
The throughput for 4, 6 and 8 PDCHs start at the same level of 14.39 kbitJs.
This is exactly the data rate needed for the chosen video sequence. The down-
292
link IP throughput per user is remaining constant as long as the necessary data
rate for Streaming is provided. Depending on the number of fixed PDCHs the
real time data rate is decreasing below the required rate of 14.39 kbitls. At
this point the IP datagram delay is increasing dramatically. With 15 users the
required data rate can not be maintained any more.
The distribution functions in Figure 6(a) and 6(b) confirm these interpreta-
tion. For one mobile station 90 percent of the IP packets for the Streaming
applications are delivered within 150 ms. The performance is not depending
on the number of PDCHs available, since the regarded mobile stations can use
only maximum 4 slots on the downlink. For 10 mobile stations and 4 PD-
CHs available more than 50 percent of the IP packets are delayed more than
300 ms and the slow increase of the distribution function indicates a high de-
lay variance, which makes the delay performance for Streaming applications
unacceptable. With 6 and 8 PDCHs 85 % of the IP packets or Streaming ap-
plications are delivered within less than 300 ms, which makes the Streaming
performance just acceptable for 10 users generating the Multimedia traffic mix.
The steps in the distribution functions are affected by the segmentation of IP
packets into radio blocks and the number of radio blocks transmitted within a
GPRS radio block period of 20 ms.
The different requirements of the applications can be supported by Quality
of Service (QoS) management functions in the RLCIMAC layer (see Stuck-
mann, 2002). The transmission of Streaming data may be privileged on the ex-
pense of background traffic. While the application response times for WWW
and e-mail would increase, the Streaming application would be able to proceed,
although high background traffic load occurred in the cell.
5. Conclusions
In this paper the performance of different Multimedia applications in packet-
switched cellular radio networks based on GPRS and EGPRS is presented. For
GPRS introduction and evolution scenarios WAP applications and a traffic mix
of WAP and conventional Internet applications over GPRS are examined. Af-
ter the performance characteristics of WAP and Internet applications have been
regarded separately, the effects of coexisting Internet traffic on WAP traffic and
vice versa are outlined. It has been shown that WAP traffic can be multiplexed
seamlessly with the Internet traffic because of the small and limited WAP deck
size, while Internet traffic slightly slows down WAP traffic in situations with
high traffic load. Regarding Video Streaming applications in coexistence with
TCP-based applications over EGPRS it has been shown that only a small num-
ber of Streaming users can be served by EGPRS, even if the percentage of
Streaming in the traffic mix is low. At least more than 4 fixed PDCHs should
be available to support Streaming applications together with background TCP
traffic. Privileged transmission of real-time data, realized by QoS manage-
ment, is one approach to provide the required bitrate for video streaming in
situations with high traffic load.
293
References
Arlitt, M. and Williamson, C. (1995). A Synthetic Workload Model for Internet Mosaic Traf-
fic. In Proc. of the 1995 Summer Computer Simulation Conference, pages 24-26, Ottawa,
Canada.
Elsen, I., Hartung, F., Hom, U., Kampmann, M., and Peters, L. (2001). Streaming Technology
in 3G Mobile Communication Systems. IEEE Computer, 34(9):46-52.
ETSI 3GPP (1998). Selection Procedures for the Choice of Radio Transmission Technologies
of the Universal Mobile Telecommunication System UMTS (UMTS 30.03, 3G TR 101
112). Technical report, European Telecommunications Standards Institute, Sophia Antipolis,
France.
Furuskar, A., Mazur, S., Milller, F., and Olofsson, H. (1999). EDGE: Enhanced Data Rates for
GSM and TDMNl36 Evolution. IEEE Personal Communications, pages 56-65.
M. Junius et al. (1993). CNCL: a C++ library for event driven simulation, statistical evalu-
ation and random number generators and distributions. Technical Report, Communication
Networks, Aachen University of Technology.
Paxson, V. (1994). Empirically-derived analytic models of wide-area tcp connections. IEEEI
ACM Transactions on Networking, 2(4):316-336.
Rose, O. and Frater, M. (1994). A comparison of models for VBR video traffic sources in B-
ISDN. IFIP Transactions C-24:Broadband communications II, pages 275-287.
Steppler, M. (1998). Performance Analysis of Communication Systems Formally Specified in
SDL. In Proc. of The First International Workshop on Simulation and Performance '98
(WaSp '98), pages 49-62.
Stuckrnann, P. (2002). Quality of Service Management in GPRS-based Radio Access Networks.
Telecommunication Systems 19:3, Kluwer Academic Publishers, pages 515-546.
Stuckrnann, P., Finck, H., and Bahls, T. (2001). A WAP Traffic Model and its Appliance for the
Performance Analysis of WAP over GPRS. In Proc. of the IEEE International Conference
on Third Generation Wireless and Beyond (3Gwireless '01), San Francisco, USA.
Stuckrnann, P. and Franke, J. (2001). The Capacity and Performance Gain Reachable with Link
Quality Control in EGPRS Networks. In Proc. of the IEEE International Conference on
Third Generation Wireless and Beyond (3Gwireless '01).
Stuckrnann, P. and Hoymann, C. (2002). Performance Evaluation of WAP-based Applications
over GPRS. In Proceedings of the IEEE International Conference on Communications (ICC
'02), New York, USA.
Stuckrnann, P. and Paul, O. (2001). Dimensioning GSM/GPRS Networks for Circuit- and Packet-
Switched Services. In Proceedings of the 10th Symposium on Wireless Personal Multimedia
Communications, ISBN 87-988568-0-4, pages 597-602, Aalborg, Denmark.
Walke, B. (2001). Mobile Radio Networks - Networking, Protocols and Traffic Performance,
2nd Edition. John Wiley & Sons, Chichester.
Willinger, W., Taqqu, M., and Sherman, R. (1997). Self-similarity through High-Variability:
Statistical Analysis of Ethernet LAN Traffic at the Source Level. 1EEEIACM Transactions
on Networking, 5:71-86.
Wireless Application Protocol Forum (1999). Wireless Application Protocol - Wireless Appli-
cation Environment Overview. Technical report.
Wireless Application Protocol Forum (2001). Wireless Application Protocol 2.0 - Wireless Ap-
plication Protocol Architecture Specification. Technical report.
Zaddach, M. and Heidtmann, K. (2001). Measurement and Traffic Characterization of H.26x-
coded Video Streams. In Proc. of the 11th GIIITG Conference on Measuring, Modelling and
Evaluation of Computer and Communication Systems. VDE-Verlag, Berlin.
VoiceNideo over IP with Multi-Class QoS in 3G
Mobile Networks
Network Technology Research Centre, School ofEEE, S2, Nanyang Technological University,
SINGAPORE 639798, elzhang@ntu.edu.sg
Key words: Video over IP, Voice over IP, QoS, queuing systems.
295
X. Lagrance and B. labbari (eds.),
Multiaccess, Mobility and Teletraffic for Wireless Communications, Volume 6, 295-308.
© 2002 Kluwer Academic Publishers.
296
1. Introduction
The services of voice over IP and video streaming over IP are going to
be introduced into the 3G mobile networks [2], since the capacity of such
network is powerful enough to handle those services. On the other hand, the
migration of IP into the 3G mobile networks is able to significantly increase
the network utilization. It also makes the connectivity of the mobile network
to be easily extended to the Internet anywhere irrespective of the user's
location. [5] [6]
on the fluid flow model. The performance evaluation focuses on the trade-off
between the delay sensitive traffic and delay insensitive traffic in terms of
traffic throughput and packet loss probability.
_I
=
480 bytes
Video payload
4bits
I Number Field
4bits
Sub SN
Packet Format for G.711 (10 ms) Clear Channel Voice 64kb/s
Voice payload
User
Header --l
Data
Sequence Priority
Flag Label Reserve Data
Number Field
Sequence Payload
Flag Label Reserve Data
Number type
80bytes
output ports. The non-blocking switching function includes that the IP packet
streams carried on the same input link are demultiplexed at the input port and
then routed to the corresponding output port according to the label assigned to
the packet stream. At the output port, IP traffic streams with different priority
levels are multiplexed before they are transmitted onto the output link. The
multiplexer at the output port consists of two parallel output queues
corresponding to two different delay priorities, respectively. Each output
queue operates on a first-in-first-out (FIFO) non-preemptive basis while the
queue is being served.
The multiple loss priority [3] is implemented using threshold control
mechanism over the partitioned buffer of the two output queues, respectively.
As shown in Figure 4, the high priority queue is fed with non-congestion
controlled traffic (i.e., delay sensitive traffic) which consists of KI classes of
packet loss priorities corresponding to the threshold Ql, Q2, .. " Qk.... QKl ,
respectively, where k is the priority index and k= 1 (i.e., Ql ) corresponds to
the highest packet loss priority. When the buffer occupancy of the high delay
priority queue exceeds the threshold Qk" ,(J < k <KI), in this case, only the
non-congestion controlled packets with loss priority ranging from 1 to k-l are
permitted to input the queue while packets with the other priority classes are
discarded. Likewise, the low delay priority output queue consisting of QKl+l,
QKl+2, ... QK2 different thresholds for packet loss control is fed with congestion
controlled traffic, where k= KI +1 corresponds to the highest packet loss
priority. Since the non-congestion controlled traffic has higher priority than
those of the congestion controlled traffic, the packets contained in the high
priority queue are always served first and the packets contained in the low
priority queue are only served when the high priority queue is empty.
QJ Q 2 High Priority Queue
l--'Ii~----------------Tltl----~
:
QD~+~JQrK~J~+2_____ __Q_u_eu_e__-r~~
Lo_w_p_r_io_rity
K1+1---+O- . -
I r---~
K2~ L.......J'---_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _--'--'----'
+ {l- [(N - i)a + ip]M }r: k[I, x - (iAk - C)~t] + 0 (~I), (1)
=I
k
where Ak Aj and Aj is the arrival rate of the traffic with j priority class
j=l
in the ON state and the term x-(iA\C)Llt is the buffer occupancy. On the right
side of equation (1), the first term is the probability of transition from the state
(i-I) to the state i at time t+ LIt, the second term is the probability of transition
301
from the state (i+ J) to the state i, the third term is the probability that the
system state i is not changing at time t, and the term o(M) represents the all
higher order terms which go to zero much rapidly than /).t when /).t intends to
zero. Hence, the effects of o(/).t) is negligible when /).t is small enough. In
equation (1), it also assumes that F.j(t,x) and FN+J(t,x) are set equal to zero.
Now, F/(t+/).t,x) and F/(t,x-t1x) are expanded for t1x =( i).k_C ) /).t in
their respective Taylor series with the assumption that the appropriate
continuity conditions are satisfied. Let /).t go to zero, the equation (1)
represents the following differential function:
By defining Fk (x) == [FOk (x), Flk (x ), ..... .F~ (x) ]r ,then the boundary
conditions in equation (2) can be obtained as below
F;Kl(O)=O if i E E uKJ ,
The arrival rate for the traffic with the k-th loss priority IS given
302
N
by A k = I)'A k ~ , (J ~ k ~ Kl). Then the packet loss probability due to
;=0
buffer overflow is given by
PLk =1- Tk / Ak, 1< k ~ Kl
(5)
Hence, the steady state distributions under the above boundary conditions is
given by
The throughput and packet loss probability for the delay insensitive traffic
with different loss priority can be calculated using equation (7).
In the second case, the high priority queue does not have any input traffic
stream. Assuming that the input streams of the high priority queue consist of j
streams which are in the ON state. We define C~ = C - P..J, where
303
Kl
Al = L A j ' Replacing F/ (x) with F/ (x), in equation (6), we have the
j=l
following differential function:
where F/
(x) is the cumulative probability distribution for the traffic with the
k-th loss priority when the system in the state i, The boundary conditions for
such a case is given by
~\O)=O l'f'
i E nriG
uj
~+l(Q)=P; l'fi E
' nrKltl
Dj
(9)
~(g)=~-l(QJ if i E l{/ U E;;j
Kl+2$k$K2
T2kj = t
;=0
Fj;-l (Qk) i}.k - L [( i}.k - C~) (Fj~-l (Qk-J) - Fj~ (Qk ))l
iEE~nEtjl J
KI+2 $k $ K2
4. Numerical Results
The following numerical results focus on the effect of priority on the
steady-state performance including the delay sensitive traffic and the delay
insensitive traffic with different packet loss priority classes, For the
illustrative purpose only, the high delay priority queue oflength Q4 consists of
2 classes of packet loss priority, named Class 1 and Class 2, where the Class 1
traffic has the higher packet loss priority than the Class 2 traffic, The
304
threshold Q2 is used for the packet loss priority control in the high delay
priority queue. Likewise, the low delay priority queue of length Q3 also
consists of 2 classes of packet loss priority, named Class 3 and Class 4, where
the Class 3 traffic has the higher packet loss priority than the Class 4 traffic.
The threshold Q4 is used for the packet loss priority control in the low delay
priority queue. The traffic of Class h (h=1,2, ... , 4) consists of Nh
homogeneous independent ON-OFF sources in which the ON state and the
OFF state are exponentially distributed, respectively, with different mean
values.
From the QoS requirement of view, video generally requires the
network to provide QoS guarantees with respect of both packet delay
and loss. Voice is sensitive to delay and delay jitter rather than packet
loss and data transfer applications are more sensitive to packet loss
rather than packet delay and delay jitter. In the following evaluation,
video and voice are classified as the Class 1 traffic and the Class 2
traffic, respectively. The Data transfer applications are classified as the
Class 3 and the Class 4 traffic in the low delay priority queue
depending on their packet loss priorities, respectively. The threshold of
Q2 and Q4 control the impact of the trade-off between delay and loss in
the higher delay priority queue and the lower delay priority queue,
respectively. However, the buffer length of Q4 and Q3 determine the
trade-off between the packet loss probability and the maximum delay.
The choice of the buffer length should be reasonably large so that the
loss probability of higher priority traffic is very small.
The effects of priority on the perfonnance of packet loss probability and
throughput for the delay sensitive traffic and the delay insensitive traffic are
illustrated in Figure 5 and Figure 6, respectively. It can be seen that the traffic
load offered by the Class 1 has the significant effects on the perfonnance of
packet loss probability for the all the other lower priority traffic classes.
Therefore, in order to achieve the desired packet loss probability for different
traffic classes, both the overall traffic load in the network and the traffic load
offered by the Class 1 need to be properly controlled. In addition, when the
traffic load of Class 1 decreases, the throughput of Class 1 traffic decreases
obviously. By contrast, the throughput of the other lower priority traffic
classes increases significantly.
Figure 7 and Figure 8 illustrate the throughput and the packet loss
probability for different threshold values, respectively, where the traffic
offered load of the Class 1 is fixed at 15% of the link capacity. Figure7
demonstrates that the effect of the different threshold values on the
perfonnance of packet loss probability is significant for all traffic classes. For
example, when the traffic loading is 80% and the threshold value of Q2 and Q4
increase from 2 to 6 respectively, the packet loss probability for the Class 2
traffic is reduced and the packet loss probability for the Class 4 is also
305
improved. Likewise, the different threshold values make the same significant
effect on performance of the throughput for the Class 3 and the Class 4 traffic.
5. Conclusion:
A novel encapsulation format for adapting voice, video-conferencing and
data traffic with multi-class QoS in the 3G mobile networks is presented. The
3G mobile IP router is modeled as a non-blocking tandem switching system
associated with output queues. The introduction of multi-class priority defined
makes the QoS control in the 3G mobile networks flexible. The multi-class of
QoS is implemented using multiple space-queues for delay priority control
and threshold controlled partial queues for loss priority control. The
performance is evaluated using fluid flow model. The illustrated numerical
results have demonstrated that the proposed incorporating priority is able to
guarantee the QoS for the transmission of VoIP and video-stream over IP over
the 3G mobile networks. However, the priority schemes do not reduce the
total packet loss but do protect the high priority traffic from packet loss while
allowing the performance of the low priority traffic to degrade as little as
possible, especially when the traffic loading and the threshold value are
properly controlled. The behavior of multi-class priority scheme is studied
with a variety of traffic conditions. The obtained results show that the high
priority traffic improve vastly with the use of multi-class priority scheme
under the condition that the proportion of high priority traffic including the
offered load and the user population must be kept to a small percentage. On
the other hand, the traffic burstiness must be also carefully controlled.
Reference:
[1) Das. S.; Misra. A.; Agrawal. P .. TeleMIP: telecommunications-enhanced mobile IP
architecture for fast intradomain mobility, IEEE Personal Communications, Volume: 7 Issue: 4,
Page(s): 50 -58, Aug. 2000
[2) Lee. W.CY.; Lee. D.J.Y., Mobile IP, Personal, Indoor and Mobile Radio Communications,
2001 12th IEEE International Symposium on , Volume: I , Page(s): 88 -92, Sept. 2001
[3) S.o.Bradner. "IPng, Internet Protocol next generation", Addison Wesley, 1996
[4) A.Elwalid. D.Mitra. "Fluid Models for the Analysis and Design of Statistical Multiplexing
with Loss Priorities on Multiple Classes of Bursty Traffic", IEEE INFOCOM' 92, 0415-0425,
May, 1992.
[5) Le Grand, G.; Horlail. E., A predictive end-to-end QoS scheme in a mobile environment,
Computers and Communications, 2001. Proceedings. Page(s): 534 -539Sixth IEEE Symposium
on ,2001
[6) Goodman, D.J., Packet reservation multiple access for local wireless communications,
IEEE Transactions on Communications, Aug. 1989
306
~~L----·--0~~~----~0~.7----~OB~----~0~.9-----
OIIered load
0.5
0.4 "---------
0.5 0.6 0.7 0.8 0.9
Offered load
10"
10-14c'>--__--+____f!l-__-+__~®_--_----<__--. .:...~-'---f'r-~"'lJ
0.55 0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95
Offered load
307
10
-2
. -~:~~=;~=~~~~~~0~
*~~~*~ ~~6~~
.p .• ~:>//
/ /// ~/ -<l-( 4%,No Priority)
-'- ( 10%,No Priority)
10-8
-+- ( 15%,No Priority)
10-12'----_~_~_ _ "_~~<::~~~~
0_55 0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95
Offered load
Figure 6: Effect of class I traffic load on packet loss prob. for four classes
0/
0.4
0.3 ~--~---~---~--~----.
OA O. O. OJ O. O.
Offered load
/ ,,¥ _ 6 - - 6 --<J.__
/ __ ....0--- '0,,-
$r /0' e
//
?,///0' .-'
~/
/
/
0.5
0.4 '------.~--~-:------~--~.--~
0.4 0.5 0.6 0.7 O.B 0.9
Offered load
r ~---<;------.<!>---~l
-------
~~---<fT--------
A-
__
t
/*/ . .-
... ¥ .----
;+ c/
*- o
1
"
/
B
10." <IJ ~ e-
0.55 0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95
Offered load
/
.0. (Threshold=2,With Priority)
.'. (Threshold=4,Wilh Priority)
.+. (Threshold=6,wllh Priority)
10-2 "----~~~~~~~~~~_~_._...L_._j__L __ ~
0.55 0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95
Offered load
Abstract: Mobile Extranets (MEs) are seen as a natural evolution of the Internet and
Intranet concepts. Their role is to provide universal (ubiquitous) access to
corporate Intranet services, along with security, location transparency, cost-
efficiency and QoS. The aim of this paper is to present and comment on an
approach for the realisation of an ME platform. GPRS and the IPv6 protocol
suite, with its integral mobility and security functionality, are key technologies
for the platform. Aspects covered in the platform presentation are the
requirements posed by various application types, the platform architecture, the
role and limitations of the technologies encompassed, and the prototype
terminal architectures. Finally, issues for further work are presented and
concluding remarks are made.
1. INTRODUCTION
It is often stated that wireless communications and the Internet are the
fastest growing businesses in the telecommunications market. A driving
1 This work is partially funded by the Commission of the European Communities, under the
Fifth Framework Program, within the 1ST (Information Societies Technology) project
MOEBIUS (IST-1999-11591: Mobile Extranet Based Integrated User Services).
309
X. Lagrance and B. Jabbari (eds.),
Multiaccess, Mobility and Teletraffic for Wireless Communications, Volume 6, 309-322.
© 2002 Kluwer Academic Publishers.
310 K.Koutsopoulos, NAlexiou, C.Konstantinopoulou, P.Demestichas,
MTheologou
force for wireless operators is the foreseen increased revenue that can come
from the (new) IP-based services. Internet Service Providers (ISPs) wish to
make their services more attractive by incorporating wireless access and
mobility capabilities. Moreover, recent times witness the explosion of the
corporate Intranet and, as a natural evolution, the Mobile Extranet concepts.
An Intranet has two fundamental functions: (a) to provide secure and
customised access to information found in transaction systems; (b) to allow
users to act on that information. Customisation refers to the user preferences,
requirements and privileges.
The aim of Mobile Extranets (MEs) is to enable ubiquitous (universal)
access to corporate Intranet services (which are otherwise geographically
limited). This access to the Intranet from the outside world should be
transparent to the applications (i.e., it should leave them unaffected), as well
as secure, cost-efficient and at the appropriate Quality of Service (QoS)
levels. In the light of these aspects, it is envisaged that the realisation of MEs
can rely on the cooperation of radio access technologies, public Internet
segments, Mobile IP (MIP) and security mechanisms. Following this trend,
the aim of this paper is to present a specific approach for the realisation of an
ME platform. The presented approach has been developed and validated in
[1]. In addition to the user equipment and the Intranet, our approach
discusses on the General Packet Radio Service (GPRS) [2], IPv6 [3], Mobile
IP (MIP) [4],[14] and IPsec [5] technologies. The paper discusses on the
role, and the potential limitations, of these technologies in the ME platform.
In this respect, the paper is on the interworkinglintegration of technologies.
This is not a trivial task. The usefulness and success of the resulting ME
platform will be a driving force for the adoption of the technologies. The
literature does not have many such studies and attempts.
The starting point for the ME platform presentation is a discussion on
high-level application types that can benefit from the ME concept (section
2). The aim of the discussion is to describe a general set of requirements that
should be met by the ME platform and to obtain insight on the added value
(additional capabilities) introduced in applications by the ME concept. In
this respect, the Integrated Mobile Health Care Solutions (IMHCS) and
Intranet Business Applications (rnA) concepts (which are the focus in [1])
will be briefly discussed.
The next, main phase of the ME presentation is the description of the
overall architecture, the presentation of the network technologies in the
platform, and the discussion on the corresponding prototype terminal
architectures and functionality (sections 3 - 5). The last phase of the ME
platform presentation focuses on sample interactions among the elements of
Establishment ofMobile Extranets through Mobile IPv6 and GPRS: 311
Enabling Universal Access to Corporate Intranets
the architecture (section 5). Finally, concluding remarks are made and
directions for future work are identified (section 7).
2. APPLICATION REQUIREMENTS
This section starts for the brief description of general application types
(IMHCS and IBA) that can benefit from the ME concept. The aim is to
identify general requirements, in terms of data transfer, mobility and
security, that the ME platform should meet.
The IMHCS concept derives when legacy health care solutions are
offered through an ME platform. The concept should enable the remote
monitoring and provision of day-to-day support to roaming patients (which
should maintain their normal lifestyle as much as possible), as well as the
detection of irregularities in the patients' condition, as far as these can be
automatically ascertained from the measured parameters. The most common
usage scenario envisages patients uploading measured data (in a transparent
manner to them), and (in return) receiving regular advice (or notifications of
critical events), either by the system itself, or by their doctors. Patient
records are stored in a database system in the Intranet, which must be
secured against intrusion of any kind, whilst simultaneously being highly
available to authorised users (e.g., doctors).
The combination of the IBA and the ME concepts enables users (e.g.,
salesmen, managers, customers, etc.) to realise remotely the transactions
offered in the Intranet limits. Sample aspects are the acquisition of
supporting material, the placement of orders, and the inspection of customer
profiles, product catalogues, price lists, order and account status, etc.
In the light of the aspects above the rest of this section summarises some
general requirements that should be met by the ME platform for properly
supporting the IMHCS and IBA application types.
Data transfer characteristics. Typically, the data types correspond to
markup text, with more demanding information flows (e.g., multimedia
presentations in the IBA case) being less likely. The information flows are
bursty, and should be provided at certain quality levels (rate, delay and
reliability requirements). These general characteristics push for the use of
flexible resource allocation schemes for achieving cost-efficiency.
Mobility and availability. Users should be able to initiate sessions from
"anywhere" within the covered geographical area (especially crucial in the
IMHCS case). Moreover, users should be reachable, in principle at all times,
312 K.Koutsopoulos, NAlexiou, C.Konstantinopoulou, P.Demestichas,
MTheologou
(a)
Appliclllw" A.p~
TCP TO.
II'><
''''' I--- '''''
..,.,
-
'''''' '''''' - '''''' - II'W
I
-
. _-----
I ';,.':~~ TqfDP
MT,.SGSN GTP. TCPlVDP, L_LnJ
,,--" '''y IP L_J..uI L_UwI ",.,.,..
(b)
Figure 1. a) Architecture of the Mobile Extranet platform. (b) High level description of
protocol stacks in ME platform
The GPRS access networks comprises three types of elements. First, the
Base Station System (BSS), which is the access element of the GPRS
network terminating the radio protocols and connecting to the core elements.
Second, the Serving GPRS Support Node (SGSN), which is the node serving
the MS; when the MS attaches to the network, the SGSN creates a record
containing information regarding mobility and security. Third, the Gateway
GPRS Support Node (GGSN), which is the node terminating all GPRS
protocols. The GGSN support edge-router specific functions and interfaces
to several ISPs or corporate networks.
The Internet Public Domain, which in principle belongs to different ISPs,
interconnects the GPRS access segment with the Intranet.
The Intranet comprises three types of entities, which are related to the
applications, the Mobile IP functionality and the security functions. The
Application Server (AS) provides the information and services offered to an
Intranet user. The firewall (FW) provides security functions, such as access
314 K.Koutsopouios, NAiexiou, C.Konstantinopouiou, P.Demestichas,
MTheoiogou
control, authentication, content security, etc. The Home Agent (HAG) is the
(MIP-related) entity of the Intranet that hosts the Home Address (HAddr)
and the Care ofAddress (CAddr) for MSs that roam outside the Intranet (and
access the services through the ME platform).
Figure 1(b) depicts the protocol stacks supporting the ME platform. It
should be observed that the platform architecture does not pose any special
requirements to the network segments. On the contrary, non-widespread
technologies, like IPv6, and associated functionality are restricted to the end-
segments, namely, the MSs and the Intranet.
This section presents the rational for selecting the network technologies
in the platform, as well as the role and limitations of the technologies.
4.1 GPRS
This subsection starts from the reasons for introducing IPv6, and its
integral MIP functionality, at some parts of the ME platform (specifically,
the MSs and the Intranet). This is explained by discussing on some
disadvantages of the approach of introducing the MIP functionality based on
IPv4. Then, some important elements of the protocol operation, especially,
in common Internet infrastructures, supporting IPv4, are provided.
The MIP functionality can be implemented on the basis of IPv4 (MIPv4)
or IPv6 (MIPv6). Typically, MIPv4 can require the presence of a Foreign
Agent for representing the MS when the available addresses in the visited
network are scarce. In that case, the Foreign Agent's address plays the role
of a globally valid CAddr for the MS. Therefore, the Foreign Agent's role is
to receive information on behalf of the MS and to direct it to the MS.
Nevertheless, a Foreign Agent is not foreseen in the current versions of the
GPRS standard, and also not associated to the Intranet. These features
suggest the investigation of the potentials of using MIPv6 in conjunction
with GPRS.
Inherent advantages of IPv6 are the large address space, and the support
for Mobile IP and security, as an integral part of the protocol suite.
Moreover, through the "6 to 4" mechanisms ([9], [10]), IPv6 enables an MS
to operate in networks that are based on IPv4. An IPv6-capable MS has a
IPv6 compatible HAddr (HAddrv6). In roaming mode the MS can discover
its network neighborhood, as well as capture the frequent changes of the
316 K.Koutsopoulos, NAlexiou, C.Konstantinopoulou, P.Demestichas,
M. Theologou
point of attachment to the network ([6], [7], [8]). In each ofthese changes an
IPv4 compatible CAddr (CAddrv4) is obtained. This address is converted
into a corresponding IPv6 compatible CAddr (CAddrv6) and registered at
the HAG, through a Binding Update Request (BUReq) message (which is
also discussed in the next subsection). This registration enables the roaming
MS to be reachable and its applications to remain unaware of the network
layer changes, which are due to mobility.
Figure 2 describes in a general manner the communication between the
MS and an AS in the Intranet.
The application at the MS sends packets with source address the
HAddrv6 and destination address the AS address. At the IPv6IIPv4 level the
following occur. At a first step, the source address is replaced by the
CAddrv6 and the HAddrv6 is indicated as an extension header. At a second
step the IPv6 message is encapsulated in an IPv4 message. The MS is
configured to use as a "6 to 4" gateway the FW of its Intranet. Therefore, the
source address of the IPv4 message is the CAddrv4, while the destination
address is that of the FW. This means that the MS specifies that the end of
the IPv4 path is the firewall of its Intranet.
.A.~icQlimr
/h611P.4 JPo611J'.4
At the other end of the communication the reverse procedure occurs. All
the messages transferred inside the Intranet are pure IPv6 messages. That is,
the Intranet is seen as an IPv6 native network. If the endpoints (MS and AS)
have been granted the permission to communicate (this aspect is discussed in
the next subsection) the firewall circulates the original IPv6 content inside
the Intranet.
Likewise, when information has to be forwarded outside of the Intranet,
the firewall, which is also a "6 to 4" gateway for the Intranet hosts,
encapsulates these inside IPv4 packets that reach their destination through
the CAddrv4 address.
Establishment ofMobile Extranets through Mobile IPv6 and GPRS: 317
Enabling Universal Access to Corporate Intranets
4.3 Security
5. TERMINALS
This section presents in more detail the architecture of the MS. Figure 3
depicts the architecture of the MS.
In brief, in its prototype version an MS can be split into the Measuring
Device (MD), which is necessary for IMHCS applications, the Terminal
Equipment (TE) and the Mobile Terminal (MT).
The TE is independent from the particular type of radio access network,
and can be a laptop PC or a Personal Digital Assistant (PDA). The MT is a
GPRS capable phone. It is envisaged that future, commercial versions of the
MS will have the TE, MT (and MD if required) as functional components of
a single hardware entity.
Figure 3(a) depicts information on the communication between the
components of the MS. The following can be noted regarding the TE - MT
and the TE - MD communication.
The TE communicates with the MT and the MD via serial connections.
The majority oflaptops and PDAs have only one built-in RS-232 port. In our
approach this port is used for the communication with the MD. In the laptop
case a second serial port card be added via the standard PC-Card expansion
slot, present in all laptops. This second port is dedicated to the
communication with the MT. Of course, an alternative is to use an Infrared
connection for the TE and MT communication. In the PDA case the second
port can be based on the, so-called, compact flash cards, present in most
devices.
Establishment ofMobile Extranets through Mobile IPv6 and GPRS: 319
Enabling Universal Access to Corporate Intranets
s"rlQI Llllk
s"rlalLlllk I"ernlillal ['llllIJIII£'1I1 (Ilifrared or (,I'IlS
\h'llfjllrili~ Del 'lce tx ruion
(.\1/1)
In) siD/from TE Ilabi/,' 1"'11111101
porta'TE) I.llplfll'. pc. "f) I los~riaJ (.111)
porlalM1)
(a)
Application
TCP
lPv6
IPv4
Relay
TE MT
(b)
Figure 3. (a) Prototype MS architecture. The MD is required for the IMHCS concept. (b)
Protocol stacks in the TE - MT arrangement.
This section presents in more detail the interactions between the elements
of the ME platform that were introduced in the previous section. Figure 4
depicts these interactions. The process consists of four main phases.
320 K.Koutsopoulos, NAlexiou, C.Konstantinopoulou, P.Demestichas,
MTheologou
7. CONCLUSIONS
MEs are seen as a natural evolution of the Internet and Intranet concepts.
The aim of this paper was to present an approach for the realisation of an
ME platform. Our approach was the following. Starting point was the
collection of general requirements posed on the platform by various
application types. Next, the overall platform architecture was presented and
details on the role of the comprised network (middleware) and terminal
technologies were introduced. In the network side the discussion focused on
GPRS and the IPv6 protocol suite, with its encompassed mobility and
security mechanisms. At the next stage a sample of the cooperation of the
platform elements was given.
The following can be identified as indicative issues for further work. The
first is the investigation of alternate technologies (e.g., in the area of
security) on which the ME platform can be based. The second is the detailed
comparison of the selected ME approach with respect to alternate concepts
such as virtual private networks (VPNs). The third is the expansion of the set
of ME software components for the purpose of better managing aspects like
the QoS levels. In this direction the possibility to negotiate on these features
with operators and ISPs will be supported. The provision of QoS levels on
this heterogeneous platform is a work area enabling the success, and
providing evidence on the viability, of the ME platform.
REFERENCES
1ST project MOEBIUS (Mobile Extranet-Based Integrated User Services) Web site,
www.ist-moebius.org, Jan. 2001
2 European Telecommunications Standardisation Institute (ETSI), "Digital cellular
telecommunication system (Phase 2+). General Packet Radio Service (GPRS). Service
description. Stage 2", GSM 03.60, v. 6.3.0, Apr. 1999
3 S.Deering, R.Hinden, "Internet Protocol version 6 (IPv6) specification", RFC 2460, Dec.
1998
4 C.Perkins, "IP mobility support", RFC 2002, Oct. 1996
5 S.Kent, R.Atkinson, "Security architecture for the Internet Protocol", RFC 2401, Nov.
1998
6 T.Narten, E.Nordmark, W.Simpson, "Neighbour discovery for IP version 6 (IPv6)", RFC
2461, Dec. 1998
7 S.Thomson, T.Narten, "IPv6 stateless address auto-configuration", RFC 2462, Dec. 1998
8 A.Conta, S.Deering, "Internet Control Message Protocol (ICMP) for the Internet Protocol
version 6 (IPv6) Specification", RFC 2463, December 1998
322 K.Koutsopoulos, N.Alexiou, C.Konstantinopoulou, P.Demestichas,
M. Theologou
Abstract: The paper presents the architecture of an UMTS terminal equipment optimised
for IP based communications and describes the traffic control mechanisms
required for supporting emerging 3G services.
Key words: IP, UMTS, Quality of Service, Traffic Control, Scheduling, PDP Context
1. INTRODUCTION
This paper presents some results of the work carried out in the European
1ST project ARROWS (Advanced Radio Resource Management for
Wireless Services) [1]. This project aims at providing advanced Radio
Resource Management (RRM) and Quality of Service (QoS) management
solutions for the support of integrated services within the context of
Universal Terrestrial Radio Access (UTRA). The project addresses packet
access, asymmetrical traffic and multimedia services, all based on IP. The
main objectives of ARROWS are: 1) to define and simulate RRM algorithms
for an efficient use of the radio resources; 2) to provide QoS bearer services
for packet switched flows at the UTRA; 3) to demonstrate the benefits of the
proposed algorithms and procedures by means of an IP based multimedia
testbed [2].
This paper is related to the third objective. The ARROWS multimedia
testbed consists of the following functional blocks: 1) an all-IP based UMTS
terminal; 2) an UTRAN (Universal Terrestrial Radio Access Network)
emulator, implementing the UMTS (Universal Mobile Telecommunications
System) radio interface and the relevant RNC (Radio Network Control)
functions; 3) a gateway implementing functions traditionally assigned to
323
X. Lagrance and B. labbari (eds.),
Multiaccess, Mobility and Teletraffic for Wireless Communications, Volume 6, 323-336.
© 2002 Kluwer Academic Publishers.
324 Manuel Ricardo, Rui Soares, Jaime Dias and Jose Ruela
SSGN (Serving GPRS Support Node) and GGSN (Gateway GPRS Support
Node); 4) a backbone IP network and its associated routers; 5) a server.
Only the first functional block is addressed - the all-IP UMTS terminal,
which supports multimedia applications and is implemented in a LINUX
based PC. Its architecture, the selected applications, the classification,
scheduling and shaping of IP flows as well as the interface with the UMTS
network interface, assumed to implement the UMTS Non-Access Stratum
(NAS) functions, are described.
The paper is organised in seven parts. Section 2 introduces the selected
multimedia applications that will run on the terminal. Section 3 discusses the
UMTS terminal architecture, which is biased towards the compatibility
between the IP and UMTS worlds from flow and QoS points of view.
Section 4 reviews the IP QoS facilities currently available in Linux and
presents a strategy for using them in an all-IP UMTS terminal. Section 5
describes the mechanism proposed for controlling the traffic. Section 6 gives
experimental results that validate this strategy. Finally, Section 7 presents
the main conclusions.
2. MULTIMEDIA APPLICATIONS
3. TERMINAL ARCHITECTURE
parameters. The mapping between RSVP QoS parameters and PDP context
QoS parameters is performed by this block.
RSVP implements RSVP (Reservation Protocol) [4] that, in ARROWS, is
used to guarantee end-to-end QoS to IP flows that traverse both the UMTS
and the IP backbone networks [5]. It can, in some circumstances, be avoided.
NAS module implements Non-Access Stratum functions, such as session
management and mobility management. It consists of two planes. On the
user plane, the module is offered as a standard LINUX network interface
(umtO, in "Figure I") and is able to exchange datagrams with the IP layer.
On the control plane, the NAS module is offered as a character device driver
(ldev/nasO, in "Figure 1"), through which messages for establishing and
terminating RABs are exchanged.
Applications
TCPIUDP
I RSVP
I
Forward I
I
I
Classifier I
I
Scheduler
~ I Mapper J
umtOI IdevlnasO
I NASModule
I
Figure 1 - UMTS Terminal Architecture
packet in
After a PDP context has been negotiated and the associated RAB
established, the terminal may start communicating. However, the terminal
may have more than one PDP context activated and more than one RAB
IP Traffic Control on UMTS Terminal Equipment 329
established, each with its own QoS parameters. It is, therefore, necessary to
direct the packets to the proper RAE, schedule the packets according to their
priorities and shape the traffic so that the flow sent to a RAE is compliant
with the QoS previously negotiated for that RAE.
5.1 Scheduling
Scheduling is usually required when there is the need to share a link with
limited bandwidth. Flows with higher priorities must be scheduled first,
taking care that flows with lower priorities do not starve. Thus, the concepts
of sharing and priority hold when dealing with scheduling. A flow with high
priority may require less bandwidth than another flow with lower priority.
Sharing is about bandwidth, priority about delay and jitter.
The services introduced in Section 2 can be ordered in ascending priority
as Email, Web Browsing, Video streaming and Videoconference. However,
a Video streaming session may require a larger bandwidth than a pure voice
seSSIOn.
The CBQ (Class Based Queuing) discipline [9] can be used to solve the
priority issue. This discipline is based on statistical scheduling and a
hierarchical tree of traffic classes. When a packet is received, it is classified
and associated to a leaf class. It is possible to associate bandwidth and a
priority to each class. The CBQ queuing discipline is delivered with two
schedulers: generic and link sharing. The generic scheduler must guarantee a
low delay to real time flows. The link sharing scheduler tries to avoid that
real time flows monopolize the use of the link.
5.2 Shaping
The rate of a flow can be regulated using shaping techniques. In this case,
the traffic passed to a RAE needs to conform to the bandwidth previously
negotiated for that RAE. The use of a CBQ class for this purpose is not
adequate, since none of its schedulers addresses this problem. Better results
can be achieved if a TBF (Token Bucket Flow) queuing discipline [10] is
associated to each leaf class.
The TBF consists of a buffer (bucket), filled with virtual pieces of
information (tokens) at a specific constant rate (token rate). An important
parameter of the bucket is its size, that is, the number of tokens it can store.
Each token in the bucket lets one incoming data octet to be sent out of the
queue and is then deleted from the bucket.
330 Manuel Ricardo, Rui Soares, Jaime Dias and Jose Ruela
Associating this algorithm with the two flows (token and data) gives
three possible scenarios: 1) the data arrives into TBF at a rate equal to the
token rate, which means that each incoming packet has a matching token and
passes the queue without delay. 2) the data arrives into TBF at a rate lower
than the token rate and therefore only some tokens are deleted when data
packets are sent out. Tokens accumulate up to the bucket size and can be
used to send data above the token rate, if this situation arises. 3) the data
arrives into TBF at a rate higher than the token rate. In this case, incoming
data can be sent out immediately, while the token bucket is not empty. The
accumulation of tokens allows short bursts of data to be passed without
delay and loss, but any lasting overload will cause packets to be constantly
dropped (or delayed).
In any case, the average data rate is bounded by the token rate.
I;~h.d·--- -----1
i r····;~·~······l
I I
___________.
filters
r--~--'r--~--'r--~--'r--~--'
I II II II
: CBQ :: CBQ :: CBQ :: CBQ
: Class :: Class : : Class :: Class
I II II II
Classification is based on the ToS value - packets with ToS values of Ox20
and OxOO are classified into the 64 and the 32 kbitls classes, respectively.
#!/binlsh
case "$1" in
start)
tc qdisc add dey ethl root handle 1: cbq bandwidth 10Mbit allot 9200 cell 16
avpkt 72 mpu 64
tc class add dey eth1 parent 1: classid 1:1 cbq bandwidth 10Mbit rate 10Mbit
avpkt 72 prio 2 allot 9200 bounded
tc class add dey ethl parent 1: 1 classid 1: 11 cbq bandwidth 10Mbit rate 64kbit
avpkt 72 prio 2 allot 9200
tc class add dey eth1 parent 1: 1 classid 1: 12 cbq bandwidth lOMbit rate 32kbit
avpkt 72 prio 2 allot 9200
tc qdisc add dey eth1 parent 1: 11 handle 11: tbflimit 20k rate 64kbit burst 1500
mtu 1500
tc qdisc add dey ethl parent 1: 12 handle 12: tbflimit 10k rate 32kbit burst 1500
mtu 1500
tc filter add dey eth1 parent 1: protocol ip prio 5 handle 1: u32 divisor 1
tc filter add dey ethl parent 1: prio 5 u32 match ip tos OxOO Oxffflowid 1: 12
tc filter add dey eth1 parent 1: prio 5 u32 match ip t05 Ox20 Oxff flowid 1: 11
stop)
tc filter del dey ethl parent 1: prio 5
tc class del dey eth1 classid 1: 12
tc class del dey ethl classid 1: 11
tc class del dey eth1 classid 1: 1
tc qdisc del dey ethl root
..
"
with a good accuracy. To obtain the highest accuracy in the generated traffic
and measurements, the processes were run with maximum priority. The
duration of the tests was always 60 seconds.
Application data is encapsulated with UDP and IP headers, which
contribute with 28 octets of overhead per IP packet. Therefore, it is
necessary to relate the data rate generated by the application with the raw IP
rate. Since it is the latter that is controlled by the TBF discipline, it will be
used when comparing the expected and the detected (measured) rates at the
destination (the same relation would hold at application level).
In the tables of results, the labels have the following means: 1) Generated
Data Rate represents a constant bit rate generated by the rude tool (data
only); 2) Expected IP Rate is the expected raw IP rate at the detection side.
This value is limited by the Token Bucket rate; 3) Detected IP Rate is the
raw IP rate measured at the destination. The latter must be compared with
the Expected IP Rate value and should not exceed it.
For a Token Bucket rate of 64 kbit/s and a datagram of 100 octets (data
size -72 octets), the results are shown in "Table I".
Table J - Packet size 100 octets, Token rate 64 kbitls
The critical data rate that corresponds to an IP rate equal to the Token
Bucket rate is 46080 bit/so When the data rate is below the critical value
(first line), the detected rate is similar to the expected one. When the data
rate is above that value (third line), the TBF controller is active; in this case
the detected rate is below the expected one. The error is -7.9%. The second
line, which corresponds to the critical value, exhibits the same behaviour.
The datagram size was then increased to 1228 octets (data size - 1200
octets) maintaining the same Token Bucket rate. The results are presented in
"Table 2".
Table 2 - Packet size 1228 octets; Token rate 64 kbitls
The critical data rate is now 62540 bitls, which means that the last two
lines correspond to an effective rate control.
The detected rates are now above the expected ones, but the errors are
small. For example, considering the last line, the error is 2.2%. It seems that
the packet size does have influence on the accuracy of the Token Bucket
Flow and that there should be a packet size that optimises its performance.
The impact of Token Bucket rates on the performance of the TBF was
evaluated, as well. Packet sizes of 100 and 1228 octets and Token Bucket
rates of 8 and 128 kbitls were tested. The results are shown in "Table 3". In
all cases the data rates are above the critical values and therefore the
expected IP rates are the corresponding Token Bucket rates.
Table 3 - Packet size 100, 1228 octets; Token rate 8, 128 kbitls
Although absolute errors increase with the token rate, relative errors are
similar. For a packet size of 100 octets, errors of -6.5%, -7.8 % and -8.0%
were obtained for data rates of 8, 64 and 128 kbit/s, respectively. For a
packet size of 1228 octets, errors of 3.8%,2.2% and 2.0% were obtained for
the same data rates. While with small packets, increasing the token rate
increases the relative error, with larger packets, increasing the token rate
decreases the relative error. For confirmation, another test, with a packet size
of 1228 octets, was carried out with a Token Bucket rate of 1 Mbitls. The
detected IP rate for this configuration was 1 013 344 bit/s, which means a
relative error of 1.33%, thus confirming the previous results.
Only one flow at a time was used in the previous tests. In a final test, two
flows classified into two different classes, one for 64 kbitls limit and the
other for a 32 kbitls limit were used. The results are shown in "Table 4".
The results for flow 2 are the same as those obtained on the first test. The
existence of more than one Queuing Discipline/Class does not seem to affect
the result. For flow 1, the relative error is -7,6 %, lower than for 64 kbitls
and higher than for 8 kbitls, in conformance with what has been previously
said about the effect of the token rate.
7. CONCLUSIONS
8. ACKNOWLEDGEMENTS
The authors wish to thank the support given by the 1ST research
programme of the European Union and their partners within the ARROWS
consortium: Universitat Politecnica de Catalunya, University of Limerick,
Telef6nica I+D and Telecom Italia Lab.
9. REFERENCES
[1] 1ST ARROWS project, http://www.arrows-ist.upc.es
[2] N.P. Magnani, F. Casadevall, A. Gelonch, S. McGrath, M. Ricardo, I. Berberana,
"Overview of the ARROWS Project", 1ST Mobile Communications Summit 2001,
Barcelona, Spain, September 9-12, 2001.
[3] 3GPP TS 23.107 V5.0.0, "QoS Concept and Architecture", April 2001.
[4] Paul White, "RSVP and Integrated Services in the Internet: A Tutorial", IEEE Comm.
Magazine, Vol. 35, No.5, May 1997, pp. 100-106.
[5] 3GPP TS 23.207 V5.0.0, "End-to-end QoS Concept and Architecture", June 2001.
[6] 3GPP TS 23.060 V5.0.0, "General Packet Radio Service (GPRS); Service description;
Stage 2", January 2002.
[7] iproute2+tc notes, http://snafu.freedom.org/linux2.2Iiproute-notes.html
[8] Linux Iproute2, http://defiant.coinet.comliproute2/
[9] Sally Floyd, "Notes on CBQ and Guaranteed Service", July1995.
[10] K. Dovrolis, M. Vedam, P. Ramanathan, "The Selection of the Token Bucket
Parameters in the IETF Guaranteed Service Class", Technical Report, Department of
ECE, University of Wisconsin-Madison, November 1997.
An Approach for Managing Networks and Services in
a Diversified Radio Environment!
Abstract: This paper builds on the assumption that in the Fourth Generation (4G)
wireless system context UMTS, MBS and Digital Broadcasting Systems
(DBS) can be three co-operating components that enable wireless access to IP-
based services. Managing the resources of this powerful infrastructure in an
aggregate manner and multi-operator scenario is a complex task. This paper
presents an approach to the overall UMTS, MBS and DBS network and
service management problem. Key points addressed are the development of
open interfaces with Service Provider mechanisms and the heterogeneous
managed infrastructure, performance monitoring, joint optimisation of the
UMTS, MBS and DBS resources in accordance with the service provider
requests and the changing with time environment conditions. The architecture
of a corresponding UMTS, MBS and DBS network and service management
platform is presented, in terms of logical blocks, their functionality,
implementation options and validation approach.
Key words: UMTS, MBS, Digital Broadcasting Systems, Network and Service
Management Platforms
I This work is partially funded by the Commission of the European Communities, under the
Fifth Framework Program, within the 1ST (Information Societies Technology) project
MONASIDRE (IST-2000-26144: Management of Networks and Services in a Diversified
Radio Environment).
337
X. Lagrance and B. labbari (eds.).
Multiaccess. Mobility and Teletraffic for Wireless Communications. Volume 6. 337-352.
© 2002 Kluwer Academic Publishers.
338 P.Demestichas, G. Vivier, G.Martinez, F. Galliano, L.Papadopoulou,
v.Stavroulaki, M. Theologou
1. INTRODUCTION
Servic...... Fiudne....rk
do_
IPN~
Paper scope
---.: - - - - - - -...----~/
Figure 1. A view of the future wireless-access world. Diverse transmission technologies offer
wireless access to IP-based services. This paper presents a corresponding network and service
management system
Figure 2 depicts the high level structure of the UMTS, MBS and DBS
management system. The management system is split in three logical blocks.
The first logical block is called managed system performance Monitoring
and Assessment and Service Provider mechanism Interworking (MASPI). In
essence, its role can be summarised in the following two general points. The
first is to capture the (changing with time) conditions that originate from the
environment (service area) of the managed UMTS, MBS and DBS
infrastructure; this is accomplished by monitoring and assessing the relevant
network and service level performance of the managed network elements
and segments. The second is to interwork with service provider mechanisms,
An Approach for Managing Networks and Services in a Diversified 341
Radio Environment
Managed System
Monitoring and Assessment & Resource Management
Service Provider Mechanism Strategies (RMS)
Interworking (MASPI)
Figure 2. High-level architecture of the UMTS, MBS and DBS management system
r - - - - - UMIS
. ;" MB~ DBS Netwo'k,
1. Virtual Network
Establishment Request
2. Virtual Network
Establishment Reauest
Managed System State
3. Virtual Network
Validation ReQUest
4b Rec:onfiguration Comma
Figure 3. Sample scenario of co-operation among the components of the UMTS, MBS and
DBS network and service management system
An Approach for Managing Networks and Services in a Diversified 343
Radio Environment
o 0 0
IP NETWORK
More specifically in this case, the joint management of the radio systems
is not straightforward since the different operators would not share strategic
information on their network. In this respect, the management system
introduced an important aspect. It presents a unified interface to the service
provider thanks to the MASPI element; from that point the dialog to get the
required QoS is handled in a manner that is transparent to the service
providers and their users. This dialog involves the management system
replicas that control each radio network. In this sense, some of the replicas
act as resource brokers, which means entities that control, offer and allocate
resources of a radio network segment. In this way all the radio systems are
involved and the appropriate radio technology for the required service is
selected. It should be noted that in the discussed case each UMD-NES
replica has a role limited to one radio technology, and that the management
system has the potential to simplify the relationship (particularly the billing
aspect) between service providers and the network operators.
The sample scenario of Figure 3 can be refined based on the discussion
above. A detailed description is omitted for brevity. Moreover, less
distributed scenarios can be defined, where for instance the management
system has the control of several (or all of the) radio systems and dialog with
the remaining ones through management system replicas.
f-·~~~=-=~~~~~-=··~··=~=~~=~~·~~·~~~=~;;~~~::~:~-:~~;;i~~;~;;;~~;~-" !
I
: i !
i, Ii ! i
! •
I I
i
II II !
I•
!'
I
! ii
II
I !
i iI !
~ ~ ! •
!L ~I
l~ ~j
i, I, Ii' !t
f i I ;
f I i
! I !
Il---- - ~M!~~l'!"!'_"!~~l~~'''''Ao~'m~J
;t ..... __..______. .- ___________ -. - --_....._-- ---- ... _.. _-- --- --- .. ~ ... . .:. _.... __MASPI
.... --..-.... ....... Component
~
Structure
.. ---- --_
,
.. _- ... _--_ ..i
parameter values that drive the radio resource control schemes that are
applied in the network, so as to satisfy the (new and established) contracts
with service providers and the environment conditions in a cost efficient
manner. Sample actions that fall in this category are the modification of the
number of connections that can be admitted per cell, the modification of
values related to power control strategies or of the transmission power of
certain base stations, etc. Such strategies may also be invoked in order to
relief hot spots, respond to network element failures, etc. Another action of
this category is the separation of radio resources to various concepts, e.g.,
virtual networks, paths within the virtual networks, service classes, new and
handover traffic, etc.
The list above can be refined in several dimensions. Each problem in the
list is an optimisation problem whose solution is integrated in the context of
the overall management tool.
MIDDLEWAIlE PlATFORM
UMD-NESPIa!fonn ~
Figure 8. Target software architecture of the UMTS, MBS and DBS network and service
management system
7. CONCLUSIONS
This paper presented an approach to the overall UMTS, MBS and DBS
network and service management problem. The starting point was the
presentation of the architecture of a corresponding platform, and its splitting
352 P.Demestichas, G. Vivier, G.Martinez, F. Galliano, L.Papadopoulou,
v.Stavroulaki, M. Theologou
into the MASPI, RMS and UMD-NES logical units. The internal structure of
these components, and implementation and validation issues were discussed.
In general terms, future work is targeted to the following aspects. First,
the completion of the ongoing implementation. Second, the realisation of
validation studies and the refinement of the architecture based on the
obtained results. Third, the dissemination of results that can prove the
efficiency of each component and of the overall platform.
REFERENCES
"IP-based mobile telecommunications networks", Special issue in the IEEE Personal
Commun., Vol. 7, No.4, Aug. 2000
2 "Next generation wireless broadband networks", Special issue in the IEEE Network, Vol.
14, No.5, Sep./Oct. 2000
3 U. Varshney, R. Vetter, "Emerging mobile and broadband wireless networks", Commun.
ofthe ACM, Vol. 43, No.6, June 2000
4 "The evolution ofTDMA to 3G", Special issue in IEEE Personal Commun., Vol. 6, No.3,
June 1999
5 "Wideband CDMA", Feature topic in IEEE Commun. Mag., Vol. 36, No.9, Sept. 1998
6 "IMT-2000: Standards effort of the ITU", Special issue on IEEE Personal Commun., Vol.
4, No.4, Aug. 1997
7 Digital Video Broadcasting Web site, www.dvb.org, Jan. 2001
8 Multimedia Car Platform, mcp.fantastic.ch, Jan. 2001
9 Broadband Radio Access for IP based Networks, www.ist-brain.org, Jan. 2001
10 Dynamic Radio for IP-Services in Vehicular Environments, www.ist-drive.org, Jan. 2001
11 "Software and DSP in radio", Feature topic in the IEEE Commun. Mag., Vol. 37, No.8,
Aug. 1999
12 "Software Radio", Special issue in IEEE Personal Commun., Vol. 6, No.4, Aug. 1999
13 "Design of broadband multiservice networks", Feature topic in the IEEE Commun. Mag.,
Vol. 36, No.5, May 1998
14 "Tools and techniques for the performance evaluation of communication networks",
Feature topic in the IEEE Commun. Mag., Vol. 36, No.8, Aug. 1998
15 E. Berruto, M. Gudmundson, R. Menolascino, W. Mohr, M. Pizarroso, "Research
activities, on UMTS radio interface, network architectures and planning", IEEE Commun.
Mag., Vol. 36, No.2, Feb. 1998
16 P. Demestichas, E. Tzifa, V. Demesticha, M. Louta, D. Karagiannis, M. Theologou,
"Issues in the design offuture mobile communications systems", In Proc. IEEE
International Conference on Communications (ICC'99), pp. 364-368, Vancouver, Canada,
June 1999
17 S. Vinoski, "COREA: Integrating diverse applications within distributed heterogeneous
environments", IEEE Commun. Mag., Vo1.35, No.2, Feb. 1997
18 "Distributed real-time systems", Feature topic in the IEEE Computer Journal, Vol. 33, No.
6, June 2000
19 R. Menolascino, P. Cullen, P. Demestichas, S. Josselin, P. Kuonen, Y. Markoulidakis, M.
Pizarroso, D. Zeghlache, "A realistic UMTS planning exercise", in Proc. 3rd ACTS Mobile
Communications Summit 98, Rhodes, Greece, June 1998