You are on page 1of 5

Performance Comparison of HARQ with Chase Combining

and Incremental Redundancy for HSDPA


PA1 Frenger: Stefan Parkvall, and Erik Da.hlman
Ericsson Research, Ericsson Radio Systems AB, SE-164 80 Stockholm, Sweden
Abstract-In this paper we compare two hybrid automatic
repeat request (HARQ) combining strategies that currently
are considered for the high speed downlink packet access
(HSDPA) evolution of WCDMA. The two HARQ combining schemes are Chase combining, where the retransmissions
are identical copies of the original transmission, and incremental redundancy (IR), where the retransmissions consist
of new parity bits from the channel encoder. We show in
this paper that the link-level performance of a HARQ typeI1 system can be significantly better with 1R compared to
Chase combining. The largest gains are obtained for high
channel-coding rates and high modulation orders. For low
modulation and coding schemes (MCSs), the link-level performance gains with IR are less significant. We further show
that in a system that uses link adaptation we can not expect
any large gains with IR as long as the link adaptation errors
are reasonable small. Furthermore, we show that on fading
channels there are situations when an IR system actually
performs poorer than a Chase combining system.
KeywordsWCDMA evolution, High Speed Downlink
Packet Access (HSDPA), Chase Combining, Incremental Redundancy (IR), Hybrid Automatic Repeat reQuest
(HARQ).

I . INTRODUCTION
As a first step in the evolution of WCDMA, a new concept denoted high speed downlink packet access (HSDPA)
is currently being developed within the 3GPP framework.
Two important design targets for the HSDPA concept are
to provide downlink peak data rates in the order of 8-10
Mbit/s for best effort packet based services a,nd to significantly reduce the downlink transmission delays.
Some importa.nt features that are introduced in HSDPA
are fast link adaptation, fast scheduling, and fast HARQ
with soft combining (i.e. type-I1 [l]).A new high speed
downlink shared channel (HS-DSCH) is introduced that is
shared in the time domain among the active users, similar to the DSCH in WCDMA of today. Instead of fast
power control, the HS-DSCH will use fast link adaptation
that adapts the size of the modulation alphabet and the
rate of the chaiinel encoder t.0 the fast channel fading. The
scheduler decides, based on e.g. the instanta.neous channel
qualities of all users, which user shall be assigned the HSDSCH channel during the upcoming tra,nsmission time interval. Furthermore, if a,n error is detected by the receiver,
the fast hybrid ARQ system ensures that the necessary retransmission is executed quickly.
Achieving low transmission delays for HSDPA is essential
in order to ensure good performance also together with
higher layer protocols e.g. TCP. It is important that the
bandwidth-delay product of the channel is in the order of
Phone: +46 0 75 70 1.52. Fax:
pal. frengeroera .ericsson.se.

+46 8 585 314 80. Email:

0-7803-7005-8/01/$10.00 0 2001 IEEE

the T C P window-size, or else it will not be possible t o


fully utilize the radio link. Furthermore, for small data
packets the slow-start beha.vior of TCP, and not the data
bandwidth of the channel, will limit the performance unless
the round trip time for T C P acknowledgments is small.
Thus, in order to benefit from the increased data rates
provided by HSDPA, reducing the transmission delays is a
key concern [a].
In the current UMTS radio access network (UTRAN) architecture the scheduling of users, selection of the transport
format (including modulation and coding parameters), and
the ARQ retransinissions are located in the radio network
controller (RNC). Since the HSDPA fast link adaptation
and fast scheduling will adapt to the fast fading of the radio
channel, it is necessary to move these functionalities closer
to the radio channel, i.e. to the Node-B (base station) instead. Also the HARQ termination point for HSDPA needs
to be located in the Node-B in order to reduce the delays
for retransmitted packets [ 3 ] .
In this paper we will compare two different packet combining strategies that are considered for HSDPA. These are
Chase combining, where each retransinissioii is identical to
the original transmission, and incremental redundancy (IR)
where each retransmission consists of new redunda.ncy bits
from the channel encoder. Obviously IR have the potential
of achieving better perforimmce compared to Chase combining. However, a HARQ system with Chase combining
will have lower complexity. The use of IR requires some additional signaling since the retransinission numbers needs
to be communicated to the receiver. Furthermore, IR requires larger receiver buffer size. The receiver buffer size
increases for each I R transmission aiid it is also necessary to
buffer soft bits instead of soft symbols in the mobile terminal (UE). Thus if IR is to be implemented for HSDPA the
complexity and cost of the system will be higher. Therefor
it is important to examine if there are any, large performance gains with IR, or if Chase combining can provide
comparable performance a.t lower cost.
In this paper we show that the link-level performance of
a HARQ type-I1 system in some cases is significantly better for IR compared to Chase combining. The largest gains
are obta,ined for high channel-coding rates a.nd high modulation orders. For low modulation aiid coding schemes
(r\irCSs), the link-level performance gains with IR are less
significant. Furthermore, in a. system using link adaptation
we can not expect any significa.iit gains with I R unless the
link adaptation errors are very large. The reason for this is
that Chase combining gives 3 dB additional signal energy
in the first retransmission a.nd with reasonably good link

1829

adaptation we will not need the additional coding gain that


can be achieved by IR. To show this effect we compare in
this paper the performance difference of Chase combining
and IR combining when the channel quality estimates are
poor. Poor link adaptation can be caused by e.g. high
Doppler shifts which makes the channel difficult to predict, or by rapid variations in the interference level. Large
Doppler will cause not only poor link adaptation, but large
errors in the receiver channel estimates as well. In this paper we therefore investigate the sensitivity towards channel estimation errors for different MCSs and we show that
high MCSs can only be used when the receiver channel estimates a.re very accurate. Therefore, in cases with high
Doppler i t is likely that only the lowest MCSs can be used,
for which the gains with IR are small. There could however be situations where poor link adaptation is caused by
large delays in the channel quality report feedback. In this
situation it is possible that the receiver could still obtain
accurate channel estimates while the transmitter is unable
to obtain accurate channel quality estimates for the link
adaptation. However, the requirements in accuracy on the
channel quality estimates in the transmitter and the channel estimates in the receiver differs, as we shall see later.
by several orders of magnitude. Rapid and unpredictable
variations in the received interference also causes poor link
adaptation. For this scenario we argue that the most important gain with an HARQ system is the diversity effect it
provides. If the interference level was high for the original
transmission it is likely that the situation will improve for
the retransmission. If the interference is constantly high
it becomes predictable and the link adaptation will then
become accurate.
Furthermore, we show that on fading channels there are
situations when ai1 IR system actually performs poorer
than a Chase combining system. This is due to the systematic turbo encoder used in WCDMA and the fact that
all systematic bits are included in the first transmission.
Therefore the retransmission when using IR consists only of
new parity bits. If the systematic bits in the first transmission are destroyed by a fading dip the receiver would benefit
more from a retransmission that includes the systematic
bits (as in Chase combining) than from a retransmission
that only contains parity bits (as in IR). An alternative
would be to use a partial IR scheme where all systematic
bits are included in each transmission but a new set of parity bits are sent in each retransmission. The perforinaiice of
a partial IR scheme will be somewhere in between the performance of the full I R and the Chase combining schemes.

AI

SF. L

Fig. 1. Block diagram of t h e simulated system

TABLE I
SlhlULATlON PARAhIETERS USED IN THIS PAPER.

I
I

Parameter
Value
Channel
3.84 x 10
Chip rate [Hz]
Transmission time interval (TTI)
0.67 ms (1 slot)
N.o. chips per TTI, Nchip
%reading factor. S F
3
N.o. multi-codes, L
N.o. bits per transport block, NTrBlk 320
24
N.o. CRC bits, ~ C R C
G
N.o. decoder tail bits
N.o. decoder iterations
18
Decoder metric
I Log-Max

~IODULATION
AND

TABLE I1
(hICSs)

CODING SCHENES

MCS #

Ktot

1
2
3
4
5
6

3
6
9
15
21
27

Ad
4
4
I 16
16

64
64

I
I

USED IN THIS PAPER

R
0.25
0.50
0.38
0.63
0.58
0.75

1
1

11. SYSTEM
MODEL
A dia.grain of the simulated system is shown in Fig. 1. In
the simulations performed, a number of transport blocks,
Ktot, of size N n B l k are concatenated, and a. CRC field
of size ~ C R Cbits is added to form an encoding block
of size Nullcoded = Ktot x N T , . B ~+~ m C R C . By letting
Nchip, S F , L , and
denote the number of chips in a
HSDPA transmission time interval: the spreading factor:
the number of multi-codes, and the modulation order: respectively, we obtain that the number of coded bits must
equal Ncoded = L x log,(M) x N,l,i,/SF. Consequently, the
rate of the turbo encoder becomes R = Nup-oded/N,-oded.
In this paper we have used N&ip = 2560, NmBlk = 320,
nxCRC = 24, S F = 4, and L = 3 in all siniulations performed. Furthermore, all simulations are performed on an
AWGN channel. The reason for this is that the channel is
a.ssuined to be constant during one transniission time interval. Using link adaptation we select the modulation and
coding scheme (MCS) based on the instantaneous channel
quality just prior to the transmission time. The results
presented here are thus valid for a range of low to moderate Doppler frequencies. The parameters used for t,he
simulations in this paper are listed in Table I. Six different
modulation a.nd coding schemes (hICS1-hlCSG) are siniulated and t,he parameters of these six MCSs are listed in
Ta.ble 11.

1830

I,

+0=5,IR
-0- (J =
+

10

-D-

5 , Chase

= 100, IR
(J= 100, Chase
(J

2
"

-5

-2

Fig. 2. Simulated slot error rate versus Zor/Ioc in dB for MCSl


(Kt,t = 3 ) . White and black markers are used for Chase combining and IR, respectively.

III I

1
i

First
Second
+ Third
-o-

IO
U
,
'

Fig. 3 .

12

14

16

[dB'

Simulated slot error rate versus lo,./Ioc in dB for h3CS6

(Kt,t = 27). White and black markers are used for Chase combining and IR: respectively.

111. NUMERICAL
RESULTS
In Fig. 2 aiid Fig. 3 the results with Chase combining
aiid IR are compared in terms of the slot. error rate versus
the ratio of the total received power (Io,.)and total interference ( I o c ) .Result,s for MCS1 (Kt,ot,= 3 ) are shown in Fig.
2 a,iid results for MCSG (I<,,,, = 27) are shown in Fig. 3.
In these two figures we see the performance of Chase coiiibiiiiiig (white markers) aiid IR (black markers) after the
second, third aiid fourth traiismissioiis (i.e. first: second,
and third retransmissions). We clearly see t,liat the gain
with IR is significant 0111~-for AICSG aiid not for hlCS1.
The gains mit.li IR coiiipared to Chase coiiibiiiiiig in

5
10
Instantaneous I,,rl I,,, [dB]

15

20

Fig. 4. Throughput versus instantaneous I,,/I,,


in dB. T h e parameter U is the standard deviation of the channel quality estimate
error.

terms of Io,./Ioc required t o achieve a slot error rate of


10% are listed for all MCSs in Ta.ble 111. From Ta.ble I11 we
conclude that IR gives significantly better link-level performance compa.red t o Cha.se combining for high modulatioiis
and coding schemes (MCS4-MCSG) aiid that only sma.11
differences are observed for lower modulation a.nd coding
schemes (MCSl-MCSS).
The transmitter will select which modulation aiid coding
scheme to use based on some cha.iiiiel qudity estimate. If
the error of this chaiiiiel quality estimate is small, then the
gains that we show in Table I11 may not be visible when
comparing the throughput of the systems with Chase coinbiiiing and IR respectively. Since in iiia.ny ca.ses only a
single retransmission is necessary also if Chase combining
is used we may not need the additional coding ga.in that
can be achieved by IR. In Fig. 4 we show the throughput
that can be achieved with IR and Chase combining, respectively. Which MCS t o use in each transmission is based on
the channel quality estiiiiate aiid is selected by coiiipariiig
with predefined switching points in a lookup table. The
lookup table was obtained by plotting t,he throughput for
each MCS individually a,iid selecting the intersection points
of the throughput curves as switching points. The chaiiiiel
quality estimate is assumed t o be normally distributed in
a logarithmic s a l e with a mean value equal to the true
channel quality aiid a standa.rd deviat.ioii of cr. Significant
gains with IR are observed for the case when the scheduler
has almost. no knowledge of the actual channel quality (i.e.
o = 100). However for smaller errors in t.he link a.daptation: t.he performance difference between Chase combining
aiid IR are much smaller (i.e. c = 5 ) .
In Fig. 5 we show the relative throughput increase in
percent that can be achieved by iiitroduciiig iiicreiiieiital
redundancy. \Ire see t.hat when the sclieduler has almost
no kiiowleclge of the channel yua1it.y ( n = 100) then the

1831

GAINWITH IR

pL
r

MCS
I.

:I

e;

-? 5

TABLE I11
CHASECOAIBINING AT SLER = 0.1

C:O~IP.\RED TO

IR gain 2]ld
Trans. [dB]
0.1
0.8
1.0
2.2
3.0
4.2

-10

I R gain 3rd
Trans. [dB]
0.2
1.o
1.0
2.7
4.0
5.4

-5
0
Instantaneous

[dB]

I R gain 4""
Trans. [dB]
0.2
1.o
1.o
2.7
4.2
6.1

IO

I o-J
Io-'
10-2
Io-'
Normalized Channel Estimation Error: a

1 o0

Fig. 6. T h e required I,,/I,,


in d B t o achieve a slot error rate of 10%
versus t h e normalized chaiiiiel estimation error a = o,"/u;.

15

Fig. 5 . Achievable throughput gain with IR compared t o Chase combining versus t h e momentary Io,./Ioc in dB. Results are shown
for channel quality estimation errors (5 of 5: 10: 20. and 100 dB.

gains with IR caii be as high as 70% increased throughput.


However for smaller errors in the chaiiiiel quality estimate
t.lie ga.ins a.re much smaller. For a < 5 there is no significant
difference between Chase combining aiid IR.
It is important to note that the results iii Fig. 5 are
obtained with perfect channel estimates in tlie receiver. In
iiiaiiy cases it is iiot reasonable t o assume tha.t tlie NodeB have very poor knowledge of tlie chaiiiiel quality while
t.lie receiver has perfect channel knowledge. Even though
t,here are scenarios: e.g. situat,ions involving soft handover,
when this assumption might be reasonable, it, is iiiore likely
t,hat. t.he error variance of tlie cliaiinel quality estimate in
t.he transiiiitt.er ailcl the channel estiiiiate in the receiver
are higlily correlat,ed most of tlie times. In Table 111 we
saw that. tlie largest gains with IR conies from the large
signal coiistellatioiis (i.e. 64 QAM). With poor cliannrl
estimates in the receiver these high coiist,ellatioiis can iiot
be used aiid t.he gains with IR will becoiiie significantly
siiialler tliaii what, we see in Fig. 5.
111 Fig. G we st.ucly the sensitivity of channel estiiiiat.ion
errors in t,lie receiver. The Ior/Ioc
required to obtaili a slot
~ r r o rrate of 10% is slion:n. versus the iioriiializrcl chaiinc~l
1832

30

10
0'
-15

-10

-5
0
5
Instantaneous IorI I,x [dB]

IO

Fig. 7. Achievable throughput gain with IR compared to Chase coiiibiniiig when using only h.ICSI-NCS3. Results are shown versus
the momentary I,, / I o c in dB for channel quality estimatiou errors CT of 5. 10. 20: and 100 dB.

estimatioii error ( a = a,'/.," with of defined as the error variance of the cha.niie1estimate and g ; clefiiied as the
variaace of t,lie channel), for AICS1-MCSG. We see that
t.lie required accuracy of t,he channel estimates varies significantly from MCS1 to MCSG. Heiice tlie highest MCSs
does not only require good climiiiel qualit,y, but. a.lso much
iiioi-e accurate channel est.iiiiates in t,he receiver.
For users on t.he cell border: or uscm with high mobility
it is reasonable to assume that only the lower NCSs can
lie used. Since we have seen that t,liere is 110 sigiiificaiit
gain with IR n-lim tlie link ;daptat,ioii works properly it is
interesting t o coinpare the rrsiilts olitaiiled n-hrii oiilj- S O ~

low MCSs can be selected. In Fig. 7 we show the gain that


can be achieved with IR if we are only allowed t o use the
three lowest ,IVICSs, i.e. MCS1-MCSS. We see t h a t the
gains with IR are only about 5% in increased throughput,
even for such large link adaptation errors g as 10 dB.
When introducing HSDPA it is desirable t o reuse as
much as possible of the existing functionality in the
WCDMA system, such as e.g. the turbo encoder. The
turbo encoder that is used in WCDNIA is a systematic
encoder. This means that the original transmission must
contain all systematic bits and, when IR is used, that the
retransmissions will contain only additional parity bits. On
a fading channel the channel quality may change from the
time of the first transmission t o the time of the retransmission. Thus, with IR we can expect some degradation if the
receiver only receives the parity bits in the retransmission
while the systematic bits in the original transmission are
lost. However, Cha.se combining where the retransmissions
are identical t o the original transmission, as well as partial
IR schemes where ail systematic bits are included in the retransmissions, are expected t o be more robust in this sense.
In Fig. 8 we examine this effect by varying the ratio of received energy in the original transmission ( E l ) and the received energy in the retransmission (&) while keeping the
total received energy (i.e. El + Ez) constant. A positive
value of y
10 logl, (El/Ez) thus means that the original
traiisinission contains more energy than the retransmission.
The curves show for different h/lCSs, the Ior/Iocrequired
t.o obtain a slot error ra.te of 10%. For Chase combining
(dashed lines) the performance is independent of y a.nd
for IR the best performance is achieved when the received
energy of the original and the re-transmissions are equal
(y = 0 dB). We see that for y = 20 dB: almost all received
energy is put on the original transmission and hence there
are almost no difference between Cha.se combining and IR
in this case. For large nega.tive values of y we can actually
see that Chase combining performs better that IR. The 5%
increased throughput that we observed for cr = 10 dB in
Fig. 7 assumed that the chaiinel did not change from the
original transmission unt,il the retransmission. However if
the channel does change (i.e. y # 0 dB): we see in Fig. 8
that the gains with IR compared to Chase coiiibiniiig will
be even snialler.

-20

-10

-5

IO

15

20

Y [dBl
in dB t h a t IS required t o obtain a slot error
Fig 8 Average I,,/I,,
rate of 10% versus y in d B y is the ratio of t h e received energy
in t h e original transmission and the retransmlssion Results are
shown for Chase combiiiing (dashed lines) and I R (sohd lines)

IR, which was studied in [4-G]', solves the problem with


non self-decodable retransmissions. The other drawbacks
that also the full IR scheme suffers reinaiiis however. Furthermore. for high MCSs oiily a small set of new parity bits
is included in the retransmission and the difference in link
level performalice for Partial IR and Chase combining is
therefore relatively small.
V . CONCLUSIONS

Since IR. implies larger memory requirements for the


mobile receivers ancl a larger amount of control signaling
compared to Chase combining, it is important that the increased complexity also results in improved performance.
In this paper we have shown that this may not be the case
for HSDPA.
REFERENCES
S. B. Wicker,

control systems for digital communication


and storage, Prentice-Hall, 1995.
3 . Peisa and A l . Meyer. .'Analytical model for T C P file transfers
over UMTS." in Proceedings 3G Wireless, 2001.
S. Parkvall; E. Dahlman. P. Frenger. P. Beming. and R.1. Persson. "The evolution of WCDhlA towards higher speed downlink
packet d a t a access." in Proceedings IEEE Vehicular Technology
Conference. Rhodes. Greece. N a y 6-9 2001.
Motorola: "Performance comparison of hybrid-ARQ schemes.:'
3 G P P input paper TSGR1#17(00)139G, 2000.
hlotorola: '.Performalice comparison of hybrid-ARQ schemes:
Additional results." 3CPP input paper TSGR1#18(01)0044.
2001.
Panasonic. '.Proposal of bitmapping for type-I11 HARQ:" 3 G P P
input paper TSC+R1#18(01)0031. 2001.

I V . DISCUSSION
In this paper we have shown that although IR gives la.rge
performance gains on t.he link level for high MCSs we may
not be able t o see these gains in a real system. Apart
from t,he cases we have studied in this pa.per we may add
that since t,he coverage area for the higher hICSs will be
much smaller than for the lower MCSs only a relatively
small percenhge of the UEs will be able t o benefit from
any event,ual gains with IR.
An alt,rrnat,iveto a full IR scheme is to use a Partial IR
scheine. where each retransmission consists of a repetition
of the systeinatic bit,s a i d a new set of parity bits. Part.ia1

-15

ETTOT

'These docunieiits are available on


Ilttp://\~w\~.3g],]~.(Jr~.

1833

the 3GPP lionir page