You are on page 1of 4

Fast Broadcasting

B. Stender, C. Huppert, and G. Richter


University of Ulm
Department of Telecommunications and Applied Information Theory
Albert-Einstein-Allee 43, 89081 Ulm, Germany
{boris.stender,carolin.huppert,gerd.richter}@uni-ulm.de

Abstract— Broadcast transmission, as proposed by Cover, is Therefore, we propose a technique that realizes broadcast
an optimal downlink transmission method to serve many users transmission gains whilst requiring only a low computational
by a single base station. Substituting or complementing the complexity. Furthermore, the inevitable signaling overhead is
fast scheduler in high speed downlink schemes, broadcasting
schemes should have a low computational complexity, as typically, kept on a low level.
decisions have to be made within milliseconds. We propose a
method for fast broadcasting with minimum rate constraints, This paper is structured as follows: In Section II the optimal
suited for transmissions over OFDM channels. Results obtained solution for broadcast transmissions over parallel channels
by computer simulations show that the novel scheme performs is shortly revisited. The novel scheme for performing fast
close to the optimum throughput.
broadcasting is described in Section III. Computer simulation
I. I NTRODUCTION results are presented in Section IV, and finally, we give some
concluding remarks in Section V.
In the downlink of modern mobile communication sys-
tems, high data transfer rates are obtained by using adaptive II. O PTIMAL P OWER A LLOCATION
techniques that cope with the fluctuations of the channel In this section, the results obtained for optimal power
conditions. One key element in such systems is the fast allocation over a family of K parallel Gaussian broadcast
scheduler, allocating resources for the users based on Quality- channels for M users are revisualized. Our description is
of-Service (QoS) requirements as well as on instantaneous mainly based on the work presented by Tse [5]. Applying
channel conditions. Covers [2] famous formula for Gaussian broadcast channels,
k
Usually, a scheduler assigns the resources exclusively to cer- the maximum rate Rm of user m in channel k, if normalized
tain users. Regarded from an information theoretic perspective, per complex signal dimension, is given by:
this is suboptimal. If broadcast transmission techniques are  
k
applied instead, notable performance gains can be obtained,
k αm Pk
depending on the underlying scenario. Therefore, we propose Rm = log 1 +  , (1)
nkm + i:π k (i)<π k (m) αik P k
to replace the fast scheduling algorithm by another algorithm
suited for performing fast broadcasting. where P k is the power allocated to channel k, αm k
is
The resources provided by Orthogonal Frequency-Division the fraction of this power used for user m’s signal, π k is
Multiplexing (OFDM) based systems, which are considered a permutation such that nkπk (1) ≤ . . . ≤ nkπk (M ) , where
exclusively throughout this paper, are sets of orthogonal carri- nkm is the effective noise. The effective noise is defined by
ers. In a multipath environment, each of these parallel channels nkm = N0 /gm k
, where N0 is twice the variance of the zero-
offers a different data transportation capability, due to the k
mean Gaussian noise and gm is the channel power gain of
underlying frequency selective fading process. In this paper, this user. These rates may be obtained by superposition code-
we consider the fading process to be static at least for the division or by dirty paper coding [1] techniques.
length of one codeword. By introducing the rate reward vector µ, all points on
The QoS requirement considered throughout this paper the boundary of the convex broadcast capacity region can be
concerns the minimum data rates per OFDM symbol that has attained by solving the following optimization problem:
to be supplied for each user. Beside this requirement, our
K

goal is to maximize the system throughput. k
max µm · Rm . (2)
k=1
Even if there already exists an optimal solution for the
described scenario, as presented in [3] and [5], a practical In a first step, the Lagrangian multiplier λ, known as the
realization according to the optimal solution is neither feasi- power price, has to be chosen such that the total power
ble for current applications nor desireable for most mobile constraint is satisfied:
environments. The reasons for this are the computational K K  µ +
  m
complexity that is quite demanding and the large signaling P = Pk = max − nkm , (3)
overhead that is necessary for the optimal solution. m λ
k=1 k=1
where the notation [x]+ = max(0, x) is used. Thus, the superposition code-division technique is used for broadcast
power P k allocated in channel k is a result of a water-filling transmissions, the receivers have to perform successive
procedure. The maximization operation is required to select interference cancellations. This process requires the decoding
the users with the highest weighted rate revenues q at the of channel codes prior to each cancellation step, which is also
operating points quite complex, especially as up to M − 1 cancellation steps
may be required. The second reason for not prefering the
k k
qm (z) = µm · rm (z) . (4) optimal solution is that thereby an outcome may be produced,
which is undesirable from a signaling point of view. As
Here, the rate revenue r is defined by using the integral many users may be target for being served simultaneously
representation of Shannons capacity formula for the AWGN within one carrier, the signaling overhead that usually is
channel: transmitted via reliable, and thus expensive, control channels

is significantly increased. Especially in mobile environments
P P
P 1 where the channel impulse responses change very fast, which
R(P ) = log 1 + = dz = r(z)dz . (5)
n 0 n+z 0
makes frequent user and power allocation updates necessary,
signaling the control information for many users results in a
Finally, to determine the fractional power allocation, the
prohibitively large overhead.
intervals with the highest weighted rate revenues have to be
identified for all users. Then, the width of these intervals, thus
It has been shown in [4] that a near optimum performance
upper limits minus lower limits, gives:
can be achieved by serving only two users in each channel,
without the need for performing a channel wise water-filling
k
αm · P k = max : (arg max qik (z) = m)− (6) procedure. This significantly reduces the overhead, as then
z i
only two user numbers and one value for power splitting have
min : (arg max qik (z) = m), 0 < z < Pk . to be signaled for each channel. Furthermore, the structure of
z i
the signaling messages can be kept much simpler, as its size
An example for an optimal power allocation is shown in
can be fixed. Similar results, but within much shorter time,
Figure 1.
can be obtained as described in the following.
    III. FAST B ROADCASTING
In this section, we present a fast method for performing
       broadcast transmissions over OFDM channels with low
signaling overhead. To reach the goal, i.e., to maximize the
system throughput whilst guaranteeing a minimum rate for
   
   all users where possible, the following approach enables a
user and power allocation with low computational complexity:

   Instead of allocating users and power in a single but com-
 

 

 
 

plex process, as required for the optimal solution, both tasks


  
 

  z are separated in the novel scheme. A simultaneous weighted


proportional fair scheduling algorithm is performed to serve


 exactly two users within each carrier. The ’good’ users Ug (k)
   that are served are those which maximize the weighted rate


z revenues at a low interference level zg , the ’bad’ users Ub (k)




that are served are those which maximize the weighted rate



revenues at a high interference level zb . Thus, the following


z
users are selected for transmission after a proper rate reward
Fig. 1. Example of an optimal power allocation. For clarity, only users that
will be served are depicted. In channel 1 user one and user three share power
vector µ has been found:
P 1 , in channel 2 only user three is served, and in channel K power P K is
k
shared between users two, one, and four. Ug (k) = arg max qm (zg ) and
m
As already stated in the introduction, this optimal
k
solution is ill-suited for most practical applications. The Ub (k) = arg max qm (zb ) , ∀k . (7)
m
first reason for this is the high computational complexity
at the transmitter. Here, finding a proper rate reward vector Note that it is possible that a user is simultaneously good
µ, e.g., by iteratively solving the non-linear optimization and bad user within the same channel, especially if its rate
problem is the most demanding task. Beside the transmitter requirement is high. It is also possible, that a user is selected
complexity, also some difficulties arise at the receiver. If the as the good one in some channels and as the bad one in some
others, even if such an outcome is not very likely and may be
obtained only for some strange settings.   
1 P
k 2 · log 1 + nk
if m ∈ Ug (k), Ub (k)
As shown in Figure 2, the selection of exactly two users R̃m = m

at certain interference levels results in some performance loss. 0 otherwise .


E.g., in the depicted example, user number two is not selected, (9)
even if it would be provided with some transmit power in the Once, the adequate users are selected, a broadcast transmis-
optimal scheme. By concentrating on only two interference sion, e.g., by signal superpositioning, is performed. For this
levels, a user selection according to (7) can be performed purpose, the carrier power is split-up between the good and
easily. Note that from now on, we will denote the (constant) the bad users. The power fraction αk is chosen in such a way
power used in each channel also by P , because of brevity that the computed transmission rate of the bad user is equal to
reasons. the design target, namely the 50% level of the highest possible
rate. Setting
Optimal values for zg and zb can be found for each scenario

by computer simulation, even if it turned out that the impact (1 − αk )P


of accurately adjusting the interference levels can increase Rbk = R̃bk = log 1 + k , (10)
nb + αk P
the throughput performance only slightly. Using zg = 0 and
zb = P is always a good choice. In general, the condition results in
0 ≤ zg < zb ≤ P must be fulfilled. nkb nkb k
αk = k − − 2−Rb , ∀k .
(11)
P · 2Rb P
q(z) Thus, all user rates still fulfill the minimum requirements
as the broadcast region is always greater than or equal to
the time-sharing region. The gain in system throughput due
to performing broadcast transmissions is completely to the
q 1 (z) avail of the good users. These rates are easily calculated by
the classical Shannon formula, obtained due to interference
cancellation of the bad users signal:
q 2 (z)

q 3 (z) αk · P
Rgk = log 1 + , ∀k . (12)
nkg
Figure 3 shows the target point of the splitting algorithm
that lies in the center of the time-sharing bound as well as the
zg zb P obtained rate-pair on the capacity boundary.
Fig. 2. User one and three are selected, as these users have the highest
weighted rate rewards at the selected interference levels.

capac
ity bo
The most complex part in the process of user and power und
allocation is to determine the rate reward vector µ. As the tim
concept based on relative rate rewards does not perfectly match e-s
ha
to the QoS concept, where the support of absolute data rates R̃b rin
gb
is demanded, e.g., to support some streaming applications, Rb ou
nd
both concepts must be brought in line. This can be done
by adjusting µ iteratively. To also guarantee a high system
throughput, the rate reward values µm are increased in our
scheme only for users that are suffering under bad conditions
in average in all channels, until their preliminary accumulated R̃g Rg
rate
Fig. 3. Throughput gain of broadcast transmission.


k
R̃m = R̃m (8) IV. S IMULATION R ESULTS
k
For our investigations, we assume that information is trans-
mitted over OFDM block-fading channels, which are modeled
fulfills the minimum rate requirements. Therefore, it is by a set of parallel degraded Gaussian broadcast channels that
initially assumed that each selected user will aim for a rate are considered to be static for the length of at least one code-
that is half the carrier capacity (50% time-sharing assumption) word. This confinement is necessary, as only then, the process
of interference cancellation and decoding can be performed simulations.
easily for broadcast transmissions utilizing superpositioning
Scen. 1 Scen. 2 Scen. 3
coding. Optimum 2091 1843 1752
To test the fast broadcasting scheme, we assume different Scheduling 1609 1554 1540
scenarios. Common to all scenarios is that M = 40 users Fast Broadcasting 2079 1810 1729
are served by a central base station, utilizing OFDM with TABLE I
K = 256 sub-carriers. For simulation, independent Rayleigh P ERFORMANCE COMPARISON FOR ALL SCENARIOS . A LL THROUGHPUT
channels are assumed for each carrier. Note, that in real VALUES ARE MEASURED IN BITS / SYMBOL .
OFDM schemes neighboring channels are correlated, which
is not considered here for simplicity reasons. The interference
levels that are used for user selection are zg = 0 and zb = P . As it can be seen from Table I, the throughput performance
of the novel fast broadcasting scheme significantly outper-
Scenario 1: There are two different groups in this scenario. forms the results achieved by scheduling and nearly reaches
The first group consists of 32 users with an average signal- optimum performance.
to-noise ratio of 20 dB, whereas the second group consists
V. C ONCLUSIONS
of 8 users with an average signal-to-noise ratio of 0 dB. As
a target, the minimum rate for all users is 25 bits per OFDM In this paper, we presented a method with low computational
symbol. complexity, to perform low overhead broadcast transmissions
over parallel broadcast channels. The reduction in complexity
Scenario 2: All users are uniformly distributed within the is a result of a separated user selection and power allocation
average SNR range of 0 dB - 20 dB. The requirement for the process.
minimum rate for each user is 15 bits per OFDM symbol. Computer simulation results showed that a near optimum
performance can be obtained by the novel scheme. In compar-
ison to a standard scheduling approach, a significantly higher
Scenario 3: In this scenario, the average user signal-to-
throughput can be notified.
noise ratios are log-normal distributed, where the mean is
set to 10 dB and the variance is set to 20 dB. Again, the ACKNOWLEDGMENTS
requirement for the minimum rate for each user is 15 bits per This work was supported by the German research
OFDM symbol. council Deutsche Forschungsgemeinschaft (DFG) under
Grant Bo 867/12, and Grant Bo 867/13-2. The authors would
For comparison purposes, the optimum power allocation like to acknowledge the DFG for their support.
strategy, and a simple scheduling strategy are applied to
all three scenarios. The strategy of the utilized scheduler is R EFERENCES
as follows: From all users that do not fulfill the minimum [1] M. H. M. Costa. Writing on dirty paper. IEEE Trans. Inform. Theory,
rate requirements, that one with the lowest accumulated rate IT-29(3):439–441, May 1983.
[2] T. M. Cover. Broadcast channels. IEEE Trans. Inform. Theory, IT-
allocates the best carrier still available. This is done until all 18(1):2–14, January 1972.
rate requirements are fulfilled. Each of the remaining carriers [3] D. Hughes-Hartog. The capacity of a degraded spectral gaussian broadcast
is allocated by the best user, following the Max C/I scheduling channel. Ph.D. thesis, 1995.
[4] B. Stender and C. Huppert. Power allocation with constraints over parallel
principle. gaussian broadcast channels. European Wireless, Athens, Greece, 2006.
The minimum rate requirements for all scenarios are such [5] D. N. Tse. Optimal power allocation over parallel gaussian broadcast
relaxed that no outage could be observed during computer channels. International Symposium on Information Theory, page 27,
1997.

You might also like