IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 5, NO.
11, NOVEMBER 2006
ARQ Error Control for Parallel Multichannel Communications
Zhihong Ding, Student Member, IEEE, and Michael Rice, Senior Member, IEEE
Abstract— This paper shows that the SW and GBN retrans-
mission protocols must be generalized when used in a multi-
channel communications system. The generalization takes the
form of packet-to-channel assignment rules. A general condition
governing the packet-to-channel assignment rule is derived and
important special cases are identified. Simulation results were
used to demonstrate that 1) packet-to-channel assignment im-
pacts channel utilization when the channels are different and
2) the optimal assignment rule produces a channel utilization
that is better than the channel utilization that results from doing
something else or nothing at all.
Index Terms— Automatic repeat request, MIMO systems.
I. I NTRODUCTION
H ISTORICALLY, automatic-repeat-request (ARQ) proto-
cols have been designed assuming temporally sequential
communication over a single channel [1], [2]. In these systems,
the transmitter sends one packet at a time over the channel.
For the stop-and-wait (SW) ARQ protocol, the transmitter
waits until it receives an positive acknowledgement (ACK)
or negative acknowledgement (NAK) from the receiver before
resuming transmission. If the round-trip delay is large enough,
then the SW ARQ protocol is inefficient and pipelined ARQ
protocols such as go-back-N (GBN) or selective repeat (SR)
are used. The SR ARQ protocol has the highest throughput,
but requires buffering at the receiver to accommodate packets
received out of order.
When ARQ protocols are used with communication over
multiple parallel channels, they must be generalized in order
to realize the full potential of parallel multi-channel commu-
nications. Multiple, parallel channels can be created in the
frequency domain by using Orthogonal Frequency-Division
Multiplexing (OFDM) Discrete Multitone (DMT) Modulation
[3]–[5], in the code domain using vector coding [6], [7], or in
space using multiple transmit antennas [8]. In data networks,
adjacent nodes may be connected by more than one link [9]. In
this case, the multiple links present multiple parallel channels
to the transmitter.
In this paper, the SW, GBN, and SR ARQ protocols are
analyzed over a communication link consisting of multiple
parallel channels with different transmission rates and different
packet error rates. The analysis leads to definitions of gener-
alized ARQ protocols that seek to improve the channel uti-
lization (a generalization of system throughput) when applied
Manuscript received March 11, 2005; revised August 10, 2005; accepted
August 26, 2005. The associate editor coordinating the review of this paper
and approving it for publication was A. Svensson. This work was supported
by the NSF under Wireless Initiative Grant CCR-9979452 and Information
Technology Research Grant CCR-0081476.
The authors are with the Department of Electrical & Computer Engineering,
459 Clyde Building, Brigham Young University, Provo, UT 84602 (email:
mdr@ee.byu.edu).
Digital Object Identifier 10.1109/TWC.2006.05168.
1536-1276/06$20.00 
Authorized licensed use limited to: Korea Advanced Inst of Science & Tech - KAIST. Downloaded on January 08,2023 at 01:38:06 UTC from IEEE Xplore. Restrictions apply.
3039
Channel 1
New
data Multi-
Channel 2 Multi-
channel
channel
data
. receiver
processor .
.
Channel M
Retransmission requests
Fig. 1. An abstraction of a communication system using multiple parallel
channels.
multiple parallel channels. The operation of ARQ in a system
that maps sequential data to multiple, parallel channels for
transmission is somewhat different than it is for single channel
systems. The operation of ARQ error control in a multichannel
system is described in Section II. In Section III, it is shown
that the generalized ARQ protocols take the form of channel
assignment rules for packets to be retransmitted. The channel
assignment rules are a function of the transmission rates and
packet error rates associated with each channel. Simulation
results are presented in Section IV that demonstrate the gains
that can be obtained by using the proper packet-to-channel
assignment rules. Conclusions are summarized in Section V.
II. ARQ E RROR C ONTROL IN A M ULTICHANNEL S YSTEM
The system model is illustrated in Figure 1. The commu-
nication link between the transmitter and receiver consists of
M parallel channels. The m-th channel is characterized by
its transmission rate Rm , measured in bits/symbol, and its
packet error rate Pm for m = 1, 2, . . . , M . We assume that
the signal-to-noise (SNR) of each channel is known. Given
a modulation type, it is usually straight-forward to compute
the bit error rate or packet error rate for a given instantaneous
SNR. All channel share a set of sequence numbers which are
used by the multichannel data processor to make the packet-
to-channel assignments. The number of retransmissions is not
restricted and the feedback channel is assumed error free.
Given a single input data stream, the multichannel data
processor divides the input data stream into packets and as-
signs a sequence number to each packet. The sequence number
preserves the sequential ordering of the packets in the input
data stream. As a consequence, when a previously transmitted
packet has to be retransmitted, it will have the lowest sequence
number in the transmit queue. The multichannel data processor
assigns the next M packets in the transmit queue to the M
parallel channels. The multichannel receiver generates ACKs
c 2006 IEEE
 3040 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 5, NO. 11, NOVEMBER 2006

and NAKs for each packet on each of the M parallel channels D be the “idle time” or round-trip delay measured in packet
and reassembles the accepted packets into a single data stream. times. Suppose an ACK is received for the packets sent on
Buffering is assumed available to handle out-of-order packet channels 1, 2, . . . , k and a NAK is received for channel k + 1.
reception. Since no buffering is provided at the receiver, the packets
Since the channels can operate at different transmission originally transmitted over channels k + 1, k + 2, . . . , M have
rates, the concept of channel utilization is used in place of to be retransmitted.
throughput as the performance measure. Channel utilization The probability that the first k packets in the parallel block
is the average information data rate over the parallel channel of M packets is successfully transmitted is
measured in bits/symbol. Note that for the single channel ⎧M
system, the channel utilization is the same as the normalized ⎪
⎪ 
⎪
throughput defined in [10] and [11]. ⎨ (1 − Pi )
⎪ k=M
PS (k) = i=1 . (1)
Chang and Yang [12], Wu, Vassiliadis, and Chung [13], and ⎪ k
⎪
⎪
Anagnostou and Protonotarios [14] investigated the throughput ⎪ (1 − Pi )Pk+1 k = M
⎩
performance of multichannel ARQ protocols where all of i=1
the parallel channels were identical (i.e., each has the same Thus, the channel utilization can be expressed as
transmission rate and packet error probability). A packet to
be retransmitted is simply assigned to the next available M k
k=1 i=1 Ri · PS (k)
channel. Which channel is used is unimportant, since all η= . (2)
the channels are the same. When the channels are different, 1+D
which channel is used for retransmission is important. This Using the matrices
behavior was first observed by Shacham in 1987 [15] in an ⎡ ⎤ ⎡ ⎤
analysis of overall resequencing delay. He noted that proper R1 PS (1)
⎢ R2 ⎥ ⎢ PS (2) ⎥
channel assignment for retransmission could have an effect on ⎢ ⎥ ⎢ ⎥
R=⎢ . ⎥ P = ⎢ . ⎥, (3)
throughput performance. Shacham and Shin [9] described and ⎣ .. ⎦ ⎣ .. ⎦
analyzed a modified SR ARQ protocol for used over parallel RM PS (M )
channels with the same transmission rate but different packet
error rates. the channel utilization may be expressed in matrix form as
The case of parallel channels with different transmission 1
characteristics is relevant to modern communication systems. η= RT VP (4)
1+D
Parallel channels with the same transmission rate, but different
packet error rates can occur in an OFDM system experiencing where RT denotes the transpose of R and V is an upper
frequency selective fading (e.g., some of the tones are suffer- triangular matrix consisting of all ones.
ing from more severe fading-induced attenuation than others). The expression (4) represents the channel utilization for a
The parallel channel point of view for DMT/OFDM was particular ordering of channels (as indicated by the position
exploited in [16], [17] to obtain improved bit allocation and of the channel transmission rates and channel packet error
bit loading algorithms. Recent results reported in [18] treated rates in the matrices R and P, respectively). Now consider a
DMT tones as parallel channels that could be “clustered” to new ordering represented by switching the order of channel
produce efficient fractional bit loading algorithms that did index i and channel index i + 1. In this ordering, the channel
not require significant trellis modifications in the receiver. transmission rates and packet error rates are summarized by
Likewise, parallel channels with the different transmission the matrices
rates, but the same bit error rate result in MIMO systems using  T
R = R1 R2 · · · Ri+1 Ri · · · RM (5)
spatio-temporal coding with power allocation assignments ⎡ ⎤
obtained using a “water-filling” solution [8]. PS (1)
⎢ P S (2) ⎥
In the next section, necessary conditions for packet-to- ⎢ ⎥
⎢ .. ⎥
channel assignment rules that improve channel utilization are ⎢ . ⎥
⎢ ⎥
derived for the SW, GBN, and SR retransmission protocols. ⎢ P (i − 2) ⎥
⎢ S ⎥
P = ⎢⎢ (1 − P 1 )(1 − P 2 ) · · · (1 − Pi−1 )Pi+1
⎥ . (6)
⎥
III. PACKET- TO -C HANNEL A SSIGNMENT RULES ⎢(1 − P1 )(1 − P2 ) · · · (1 − Pi−1 )(1 − Pi+1 )Pi ⎥
⎢ ⎥
⎢ PS (i + 1) ⎥
A mathematical expression for the channel utilization is ⎢ ⎥
⎢ .. ⎥
derived for each of the three retransmission protocols. This ⎣ . ⎦
expression is then used as the basis for defining packet-to- PS (M )
channel assignment rules that improve the channel utilization
for each case. The channel utilization for this ordering of channels is
1
A. Packet-to-Channel Assignment Rule for SW ARQ η = RT VP . (7)
1+D
In the SW ARQ protocol, the transmitter sends a block The difference between the two channel utilizations is
of M packets to the receiver and waits for acknowledgement
T
from the receiver before it sends the next block of packets. Let (η − η  )(1 + D) = (R − R ) VP + RT V (P − P ) (8)
Authorized licensed use limited to: Korea Advanced Inst of Science & Tech - KAIST. Downloaded on January 08,2023 at 01:38:06 UTC from IEEE Xplore. Restrictions apply.
 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 5, NO. 11, NOVEMBER 2006 3041

 
= 0 · · · 0 Ri − Ri+1 Ri+1 − Ri 0 · · · 0 VP bit error rate1 . Let Pi = LRi for i = 1, 2, · · · , M . Then
⎡ ⎤ the necessary condition (12) becomes
0
⎢ .. ⎥
⎢ . ⎥ Δi = (Ri − Ri+1 )(1 − Pi )LPi+1 − Ri+1 L(Ri − Ri+1 )
⎢ ⎥
⎢ 0 ⎥
⎢ i−1 ⎥ = L(Ri − Ri+1 )[(1 − Pi )Pi+1 − Pi+1 ]
⎢(P − P ) (1 − P ) ⎥
T
+R V⎢⎢ i i+1
i−1 n=1 n ⎥
= L(Ri − Ri+1 )[−Pi Pi+1 ] > 0 ⇒ Ri < Ri+1 .
⎥ (9)
⎢(Pi+1 − Pi ) n=1 (1 − Pn )⎥
⎢ 0 ⎥ (15)
⎢ ⎥
⎢ . ⎥
⎣ .. ⎦ This means the channels should be ordered from lowest
0 transmission rate (lowest packet error rate) to highest
transmission rate (highest packet error rate).
• All channels have different transmission rates and dif-
i−1
 ferent packet error rates but all packet error rates satisfy
= (Ri − Ri+1 ) (1 − Pn )(1 − Pi )Pi+1 Pi  1 for i = 1, 2, · · · , M . In this case, the necessary
n=1
i−1 i−1
condition (12) implies that
 
+ Rn (1 − Pn )(Pi − Pi+1 )
Ri+1 (Pi − Pi+1 )
 i−1
n=1 n=1
 i−1 (Ri − Ri+1 ) Pi+1 >
1 − Pi
 
− Rn + Ri+1 (1 − Pn )(Pi − Pi+1 ) (10) ≈ Ri+1 (Pi − Pi+1 ) (16)
n=1 n=1 Ri Ri+1
⇒ > . (17)
Pi Pi+1
 
= (Ri − Ri+1 ) (1 − Pi ) Pi+1 − Ri+1 (Pi − Pi+1 ) The interpretation of this result is that the channels should
i−1
 be ordered (in descending order) based on the ratio
× (1 − Pn ) . (11) of transmission rate to packet error rate. Note that the
n=1 second and third special cases are special cases of this
scenario.
A necessary condition for the original ordering to be optimal
is that In all cases, the ordering is based on a quantitative measure
of the channel quality. A channel with a higher transmission
Δi = (Ri − Ri+1 ) (1 − Pi ) Pi+1 − Ri+1 (Pi − Pi+1 ) > 0 rate, or lower packet error rate (or both) has a higher quality.
(12) The reason the higher quality channels should be ordered
for i = 1, 2, . . . , M − 1. Five important special cases should first lies in the details of how the SW ARQ protocol assigns
be noted. sequentially available packets in the transmit queue to the
• All channels have identical transmission rates and packet parallel channels. If a transmission in the first channel fails,
error rates. In this case, Ri = Ri+1 and Pi = Pi+1 for all then the packets sent in channels 2 and higher must also be
i = 1, 2, · · · , M − 1. Then Δi is zero for all i. Channel retransmitted. This must be the case since SW ARQ does not
ordering in the assignment rule does not matter. provide any buffering at the receiver for reordering packets
• All channels have the same transmission rate but dif- received out of order.
ferent packet error rates. Let Ri = R be the common
transmission rate for i = 1, 2, . . . , M . Then the necessary
condition (12) becomes B. Packet-to-Channel Assignment Rule for GBN ARQ
Δi = −R(Pi − Pi+1 ) > 0 ⇒ Pi < Pi+1 . (13) In the GBN ARQ protocol, the transmitter sends packets
to the receiver continuously and does not wait for acknowl-
This means the channels should be ordered from lowest edgements from the receiver. The acknowledgement for each
packet error rate to highest packet error rate. transmission block arrives after a round-trip delay of N × M
• All channels have different transmission rates but the packets (e.g., N blocks of M packets). During this interval,
same packet error rates. Let P = Pi be the common N − 1 blocks of M packets have also been transmitted. When
packet error rate for i = 1, 2, . . . , M . Then the necessary a NAK is received for a particular packet, all the subsequent
condition (12) becomes packets in the block, together will all packets in the subsequent
Δi = (Ri − Ri+1 )(1 − P )P > 0 ⇒ Ri > Ri+1 . (14) N −1 blocks, are discarded by the receiver and must be resent
by the transmitter.
This means the channels should be ordered from highest
transmission rate to lowest transmission rate. 1 To see that this is so, let P be the common bit error rate for each channel
b
• All channels have different transmission rates and dif- and suppose that the packet length, L (measured in symbols), is also the same
ferent packet error rates but the packet error rate of each for each channel. The number of bits transmitted in a length-L packet over
channel m with rate Rm is Lb,m = LRm . The packet error rate may be
channel is proportional to the transmission rate. This case expressed as Pm = 1 − (1 − Pb )Lb,m ≈ Lb,m Pb . Substituting we obtain
occurs when each channel is designed to have the same Pm ≈ LPb Rm which may be expressed as LRm . using L = LPb .
Authorized licensed use limited to: Korea Advanced Inst of Science & Tech - KAIST. Downloaded on January 08,2023 at 01:38:06 UTC from IEEE Xplore. Restrictions apply.
 3042 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 5, NO. 11, NOVEMBER 2006

5
Let S be the average number of accepted blocks prior to a r =r =2.0 (Analytical Result)
P R
NAK. S may be expressed as r =r =2.0 (Simulation Result)
4.5 P R

∞
M k  M
 r =r =1.4 (Analytical Result)
   P R
rP=rR=1.4 (Simulation Result)
S= k· (1 − Pi ) 1− (1 − Pi ) 4
r =r =1.1 (Analytical Result)
P R
k=1 i=1 i=1 r =r =1.1 (Simulation Result)

Channel Utilization
∞
 3.5
P R

= [1 − PS (M )] k · PSk (M )
k=1 3
PS (M )
= . (18)
1 − PS (M ) 2.5

The channel utilization for GBN ARQ is

2
M M−1
 k 
S  1   PS (k)
η= Ri + Ri . 1.5
N + S i=1 N +S i=1
1 − PS (M )
k=1
(19) 1
Using the matrices 0 0.02 0.04 0.06 0.08 0.1
⎡ ⎤ ⎡ ⎤ P
1
R1 PS (1)
⎢ R2 ⎥ ⎢ PS (2) ⎥ 11
⎢ ⎥ ⎢ ⎥
R̄ = ⎢ . ⎥ P̄ = ⎢ .. ⎥, (20) rP=rR=2.0 (Analytical Result)
⎣ .. ⎦ ⎣ . ⎦ 10 rP=rR=2.0 (Simulation Result)
RM−1 PS (M − 1) rP=rR=1.4 (Analytical Result)
9 rP=rR=1.4 (Simulation Result)
the channel utilization may be expressed in matrix form as rP=rR=1.1 (Analytical Result)
8
M rP=rR=1.1 (Simulation Result)

Channel Utilization

S 1 1
η= Ri + · R̄T V̄P̄ (21) 7
S+N i=1
N + S 1 − P S (M )
6
where V̄ is an upper triangular matrix consisting of all ones.
5
Note that the first term in (21) does not change with
re-ordering of the channel index. The matrix form of the 4
second term is also similar to the matrix form of (4) for
SW-ARQ. Applying the same line of reasoning to this case 3
produces the same necessary condition (12) for improved 2
channel utilization.
1
0 0.02 0.04 0.06 0.08 0.1
C. Packet-to-Channel Assignment Rule for SR ARQ P1

In the SR ARQ multichannel protocol, the transmitter 16

sends packets to the receiver continuously and re-sends only rP=rR=2.0 (Analytical Result)

those packets that were negatively acknowledged. Assuming rP=rR=2.0 (Simulation Result)

a sufficiently large buffer at the receiver to reassemble the 14 rP=rR=1.4 (Analytical Result)

packets received out of order, the channel utilization can be r =r =1.4 (Simulation Result)
P R

expressed as rP=rR=1.1 (Analytical Result)

12
Channel Utilization

M r =r =1.1 (Simulation Result)

P R
η= Ri · (1 − Pi ). (22)
i=1
10
Reordering the channel indexes in (22) does not effect the
channel utilization. Thus, channel utilization is independent
8
of the channel assignment for the SR ARQ protocol.

IV. S IMULATION R ESULTS 6

Computer simulations were used to assess the accuracy of
the throughput expressions derived in Section III. For our 4
numerical example, we consider a 4 channel system (i.e. 0 0.02 0.04 0.06 0.08 0.1
P1
M = 4) where the idle time of SW-ARQ is D = 2 block
times and the round-trip delay of the GBN-ARQ is N =
Fig. 2. Comparison of computer simulations with analytical expressions for
3 block times. To illustrate the effect of different channel channel utilization for the three ARQ protocols: (top) stop-and-wait; (middle)
characteristics on the channel utilization, we adopt the same go-back-N ; (bottom) selective-repeat. In all cases, R1 = 1 bits/symbol and
technique used in [9]: Pi+1 /Pi is a constant (which we call rP = rR to simplify the presentation.

Authorized licensed use limited to: Korea Advanced Inst of Science & Tech - KAIST. Downloaded on January 08,2023 at 01:38:06 UTC from IEEE Xplore. Restrictions apply.
 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 5, NO. 11, NOVEMBER 2006 3043

50
where ν is the vector of noise samples and and Z the vector
Pre−ARQ BER ≤ 5×10−3
45 of information symbols. In this way, spatio-temporal coding
Pre−ARQ BER ≤ 2×10−3
creates rank(H) parallel communication channels. The gains
40 Pre−ARQ BER ≤ 10−3 of each of the channels is given by its singular values λn which
35 is the element (n, n) in the matrix Λ. Different information
rates are assigned to each of the sub-channels using a spatio-
30 temporal water-filling solution to achieve capacity.
Gain (%)

25 Simulations were performed using two 4 × 4 channel

matrices: the first was the IID MIMO channel where H
20 consists of 16 zero-mean unit-variance complex Gaussian
random variables. The second channel was measured in an
15
indoor environment as described in [20]. The channel matrix
10 is assumed constant during one packet interval but varies from
packet to packet.
5
A set of Gray-coded M-PSK modulation schemes for M =
0 2, 4, 8, 16, 32, 64 was used to provide transmission rates of 1,
50 100 150 200 250
2, 3, 4, 5, and 6 bits/symbol, respectively. The binary reflected
Packet Length (symbols)
Gray code described in [21] was used for the bit-to-symbol
50 mapping. The symbols were indexed 0, 1, . . . , M − 1 starting
−3
Pre−ARQ BER ≤ 5×10 with the point 1 + j0 and proceeding in the counter-clockwise
45 −3
Pre−ARQ BER ≤ 2×10 direction. Since this set of modulation schemes provide finite
−3
40 Pre−ARQ BER ≤ 10 granularity in the transmission rates, a bit loading algorithm
[3] is used to allocate power and bit to the parallel channels
35
subject to a pre-ARQ bit error rate constraint.
30 The channels are ordered λ1 ≥ λ2 · · · ≥ λr so that the
Gain (%)

transmission rates of the r sub-channels satisfy R1 ≥ R2 · · · ≥

25
20
15
10
5 using that assignment rule was simulated and designated
0
50 100 150 200 250
Packet Length (symbols)
Fig. 3. Performance comparison for the multichannel SW-ARQ retransmis-
sion protocol for a 4 × 4 MIMO channel using spatio-temporal coding: (top)
IID channel; (bottom) BYU channel.
rP ) for i = 1, 2, . . . , M − 1; Ri+1 /Ri is a constant rR for
i = 1, 2, . . . , M − 1.
The results are summarized in Figure 2 where we see that
the simulation results matched the analytical expressions ex-
actly. A comparison of the packet-to-channel assignment rules
derived in Section III to other packet-to-channel assignment
rules as a function of rP and rR is plotted in [19].
Simulations were also performed for a MIMO system with 4
transmit antennas and 4 receive antennas using spatio-temporal
coding [8]. Spatio-temporal coding over a frequency non-
selective MIMO channel H is accomplished by computing
the singular value decomposition of the channel matrix: H =
UΛV∗ and using the right singular vectors (columns of V)
as the bases for the transmitted sequences and the left singular
vectors (columns of U) as the matched filters. The vector of
matched filter outputs may be expressed as
R = U∗ HVZ + ν = ΛZ + ν (23)
Authorized licensed use limited to: Korea Advanced Inst of Science & Tech - KAIST. Downloaded on January 08,2023 at 01:38:06 UTC from IEEE Xplore. Restrictions apply.
Rr while the packet error rate of the r channels satisfy
P1 ≥ P2 · · · ≥ Pr . Thus, this scenario matches the forth
special case in Section III. The packet-to-channel assignment
rule developed in Section III requires the packets to be ordered
from lowest transmission rate to highest transmission rate (that
is, in the reverse order from above). The channel utilization
ηopt . To model the effect of “doing nothing,” the channel
ordering λ1 ≥ λ2 · · · ≥ λr was left in place and packets
to be retransmitted were inserted in the next available channel
without regard for the channel number. The channel utilization
for this case was also simulated and designated ηstatic .
The simulation results are summarized in Figures 3 and 4
for the SW ARQ protocol (with idle time D = 2) and the GBN
ARQ protocol (with N = 3), respectively. In these plots, the
channel utilization gain
ηopt − ηstatic
G= (24)
ηstatic
is plotted as a function of packet length. Observe that sub-
stantial gains in channel utilization can be realized, especially
as the packet length increases and the pre-ARQ bit error rate
increases.
V. C ONCLUSION
In this paper, we have shown that the SW and GBN
retransmission protocols must be generalized when used in
a multichannel communications system. The generalization
takes the form of packet-to-channel assignment rules. A
general condition governing the packet-to-channel assignment
rule was derived and important special cases were pointed
out. Simulation results were used to demonstrate the that
1) packet-to-channel assignment impacts channel utilization
 3044 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 5, NO. 11, NOVEMBER 2006

50
R EFERENCES
Pre−ARQ BER ≤ 5×10−3
45 [1] S. Wicker, Error Control Systems for Digital Communication and
Pre−ARQ BER ≤ 2×10−3
Storage. Upper Saddle River, NJ: Prentice-Hall, 1995.
−3
40 Pre−ARQ BER ≤ 10 [2] J. Shu Lin, Daniel J. Costello, Error Control Coding, Fundamentals and
Applications. Upper Saddle River, NJ: Prentice-Hall, Inc., 1983.
35 [3] J. Bingham, “Multicarrier modulation for data transmission: an idea
whose time has come,” IEEE Commun. Mag., vol. 28, no. 5, pp. 5–14,
30 May 1990.
Gain (%)

[4] B. Saltzberg, “Performance of an efficient parallel data transmission

25 system,” IEEE Trans. Commun. Technol., vol. 15, pp. 805–811, Dec.
1967.
20 [5] J. C. Tu, “Theory, design and application of multi-channel modulation
for digital communications,” Ph.D. dissertation, Stanford Univ., Stan-
15 ford, CA, 1991.
[6] J. Lechleider, “The feasibility of using adaptive transmitters to suppress
10 crosstalk,” in Proc. IEEE International Conference on Communications
1989, vol. 1, pp. 548–551.
5 [7] S. Kasturia, J. Aslanis, and J. Cioffi, “Vector coding for partial-response
channels,” IEEE Trans. Inf. Theory, vol. 36, pp. 741–762, July 1990.
0 [8] G. G. Raleigh and J. M. Cioffi, “Spatio-temporal coding for wireless
50 100 150 200 250 communication,” IEEE Trans. Commun., vol. 46, pp. 357–366, Mar.
Packet Length (symbols) 1998.
[9] N. Shacham and B. C. Shin, “A selective-repeat-ARQ protocol for
50 parallel channels and its resequencing analysis,” IEEE Trans. Commun.,
−3
Pre−ARQ BER ≤ 5×10 vol. 40, pp. 773–782, Apr. 1992.
45 −3 [10] J. Yun, W. Jeong, and M. Kavehrad, “Throughput analysis of selective
Pre−ARQ BER ≤ 2×10
repeat ARQ combined with adaptive modulation for fading channels,”
−3
40 Pre−ARQ BER ≤ 10 in Proc. MILCOM 2002, vol. 1, pp. 710–714.
[11] J. Yun, W. Jeong, and M. Kavehrad, “Throughput performance analysis
35 for automatic-repeat-request techniques as combined with adaptive rate
transmission,” in Proc. Radio and Wireless Conference, 2002, pp. 11–14.
30 [12] J.-F. Chang and T.-H. Yang, “Multichannel ARQ protocols,” IEEE
Gain (%)

Trans. Commun., vol. 41, pp. 592–598, Apr. 1993.

25 [13] W.-C. Wu, S. Vassiliadis, and T.-Y. Chung, “Performance analysis of
multi-channel ARQ protocols,” in Proc. 36th Midwest Symposium on
20 Circuits and Systems 1993, vol. 2, pp. 1328–1331.
[14] M. Anagnostou and E. Protonotarios, “A selective repeat ARQ protocol
15 with n channels,” in Proc. GLOBECOM 1984, pp. 1445–1449.
[15] N. Shacham, “Packet resequencing in reliable transmission over parallel
10 channels,” in Proc. International Conference on Commununications
1987, pp. 557–562.
5 [16] R. V. Sonalkar, “Bit- and power-allocation algorithm for symmetric
operation of DMT-based DSL modems,” IEEE Trans. Commun., vol. 50,
0 pp. 902–906, June 2002.
50 100 150 200 250 [17] D. Dardari, “A uniform power and constellation size bit-loading scheme
Packet Length (symbols) for OFDM based WLAN systems,” in Proc. IEEE International Sym-
posium on Personal, Indoor and Mobile Radio Communications 2004,
Fig. 4. Performance comparison for the multichannel GBN-ARQ retrans- vol. 2, pp. 1220–1224.
mission protocol for a 4 × 4 MIMO channel using spatio-temporal coding: [18] L. Pierrugues, O. Moreno, P. Duvaut, and F. Ouyang, “DMT perfor-
(top) IID channel; (bottom) BYU channel. mance improvement based on clustering modulation, applications to
ADSL,” in Proc. IEEE International Conference on Communications
2004, vol. 5, pp. 2807–2811.
[19] Z. Ding and M. Rice, “Throughput analysis of ARQ protocols for
parallel multichannel communications,” in Proc. IEEE Global Commu-
nications Conference 2005, pp. 1279-1283.
when the channels are different and 2) the optimal assignment
[20] J. Wallace, M. Jensen, L. Swindlehurst, and B. Jeffs, “Experi-
rule produces a channel utilization that is better than the mental characterization of the MIMO wireless channel: data ac-
channel utilization that results from doing nothing. The gain quisition, analysis, and modeling,” IEEE Trans. Wireless Commun.,
increases as the difference between the channels becomes vol. 2, pp. 335–343, Mar. 2003. (Data sets are available on-line at
http://www.ee.byu.edu/wireless/data/probe narrow.cgi)
more pronounced. This is illustrated by the fact that the gain [21] J. Lassing, E. Ström, E. Agrell, and T. Ottosson, “Computation of
increases as the packet length (symbols/packet) increases, or the exact bit-error rate of coherent M -ary PSK with Gray code bit
as pre-ARQ BER increases. mapping,” IEEE Trans. Commun., vol. 51, pp. 1758–1760, Nov. 2003.

Authorized licensed use limited to: Korea Advanced Inst of Science & Tech - KAIST. Downloaded on January 08,2023 at 01:38:06 UTC from IEEE Xplore. Restrictions apply.