You are on page 1of 3

950 IEEE COMMUNICATIONS LETTERS, VOL. 13, NO.

12, DECEMBER 2009

An Efficient Dynamic Schedule for


Layered Belief-Propagation Decoding of LDPC Codes
Guojun Han and Xingcheng Liu, Member, IEEE

Abstract—An efficient dynamic schedule for layered belief- requires significantly less iterations, and yields better error
propagation (LBP) decoding of low-density parity-check (LDPC) performance than the standard sequential schedules. However,
codes is presented, in which the check nodes connecting to residual computation for the NWRBP greatly increases the
a variable node with maximum relative message residual are
chosen in each dynamic decoding iteration, and then LBP is decoding complexity in each iteration.
performed for these check nodes. In this schedule we combined In this letter, based on the Layered Belief-Propagation
the features of lazy schedule and node-wise residual belief propa- (LBP) [3], we designed an efficient dynamic schedule, referred
gation (NWRBP) together while keeping the extra computational to as the efficient dynamic schedule for LBP (EDS-LBP),
complexity very low. Simulation results show that the new which is derived from the lazy schedule and the NWRBP
schedule speeds up the convergence rate and greatly improves
error performance at medium to high signal-to-noise ratio (SNR) while keeping the extra computational complexity much lower
when compared to the standard LBP decoding algorithm. when compared to the NWRBP. The proposed schedule speeds
up the convergence rate and improves the error performance
Index Terms—LDPC codes, dynamic schedule, layered belief-
propagation (LBP), residual. significantly at medium to high signal-to-noise ratio (SNR) as
compared with the standard LBP.

I. I NTRODUCTION
II. BACKGROUND

L DPC codes can approach the Shannon limit under low-


complexity iterative belief-propagation (BP), or message-
passing, algorithm over the factor graph of the code. The
A binary [N, M] LDPC code graph is a bi-partite one com-
posed by N variable nodes 𝑣𝑗 that correspond to the encoded
message passing schedule over the factor graph directly affects bits, and M check nodes 𝑐𝑖 that correspond to the parity-
convergence rate, complexity and error performance. The check equations. At the beginning of the decoding process,
standard decoding schedules include the flooding schedule [1], each variable node 𝑣𝑗 is initialized with the corresponding
𝑝 (𝑦 ∣𝑣 =1)
or the simultaneous schedule, in which all the variable nodes, soft output 𝐶𝑣𝑗 = 𝑙𝑜𝑔( 𝑝𝑗𝑗 (𝑦𝑗𝑗 ∣𝑣𝑗𝑗 =0) ) from the channel detector,
and subsequently all the check nodes, pass new messages where 𝑦𝑗 is the received signal. In the BP the messages from
to their neighbors in each iteration. Besides the flooding the variable nodes to the check nodes are propagated with:
schedule, some different types of non-flooding, or sequential ∑
𝑚𝑣𝑗 →𝑐𝑖 = 𝑚𝑐𝑎 →𝑣𝑗 + 𝐶𝑣𝑗 (1)
schedules for iterative BP decoding were proposed earlier,
𝑐𝑎 ∈𝑁 (𝑣𝑗 )∖𝑐𝑖
such as check-node updating schedules ([2]-[4]) and variable-
node updating schedules [5][6]. Studies [5][7] showed that where 𝑁 (𝑣𝑗 ) ∖ 𝑐𝑖 denotes the neighbors of variable node 𝑣𝑗
sequential schedules not only speed up the convergence rate excluding check node 𝑐𝑖 . The messages propagated from the
in terms of number of iterations but also outperform the check nodes to the variable nodes are:
traditional flooding schedule for a large number of iterations. ( ∏ ( 𝑚𝑣𝑏 →𝑐𝑖 ))
For sequential updating schedule, to find the ordering of 𝑚𝑐𝑖 →𝑣𝑗 = 2𝑎𝑟𝑐𝑡𝑎𝑛ℎ 𝑡𝑎𝑛ℎ (2)
2
message update that results in the best convergence speed 𝑣𝑏 ∈𝑁 (𝑐𝑖 )∖𝑣𝑗

and/or error performance is a new problem. In order to reduce where 𝑁 (𝑐𝑖 ) ∖ 𝑣𝑗 denotes the neighbors of check node 𝑐𝑖
the number of iterations needed to converge, Levin et al. [8] excluding variable node 𝑣𝑗 . The soft decoding result for each
presented a lazy schedule that only the subset of variable variable node 𝑣𝑗 is obtained as follows:
nodes is updated in each iteration. Recently, the informed ∑
dynamic schedules (IDS) were presented [9][10], in which 𝑚𝑣𝑗 = 𝑚𝑐𝑎 →𝑣𝑗 + 𝐶𝑣𝑗 (3)
the differences between the values of the message before and 𝑐𝑎 ∈𝑁 (𝑣𝑗 )
after an update are used to dynamically update the schedule.
Simulation results show that one of these dynamic schedules, The sequential message-passing schedules, such as LBP,
called the node-wise residual belief propagation (NWRBP), update check nodes sequentially according to a predetermined
schedule. The lazy schedule that is derived from the variable-
Manuscript received July 26, 2009. The associate editor coordinating the node sequential updating schedule updates the variable node
review of this letter and approving it for publication was S. Yousefi. message with probability 𝛽(𝑙, 𝑟𝑣 , 𝑡𝑣 ), where 𝑙 is the iteration
G. Han is with the Department of Electronic and Communications Engi-
neering, Sun Yat-sen University, and with Guangdong University of Technol- number, 𝑟𝑣 is some reliability measure of variable node 𝑣 and
ogy, Guangzhou, 510275, P.R.China (e-mail: gjhan@gdut.edu.cn). 𝑡𝑣 is the number of iterations that have passed since variable
X. Liu (corresponding author) is with the Department of Electronic and node 𝑣 was updated the last time [8]. Under the same decoding
Communications Engineering, Sun Yat-sen University, Guangzhou, 510275,
P.R.China (e-mail: isslxc@mail.sysu.edu.cn). complexity, the lazy schedule speeds up the convergence when
Digital Object Identifier 10.1109/LCOMM.2009.12.091555 compared to the standard sequential updating schedule.
1089-7798/09$25.00 ⃝
c 2009 IEEE
HAN and LIU: AN EFFICIENT DYNAMIC SCHEDULE FOR LAYERED BELIEF-PROPAGATION DECODING OF LDPC CODES 951

The residual belief propagation (RBP) is an IDS strat- 0


10
Flooding
egy that schedules message updates according to the value LBP
of the message residual [9]. The residual is defined to be −1
10 EDS−LBP
NWRBP
the difference between the values of check-to-variable node
message before and after an update. The RBP updates first −2
10
the message with maximum residual. The decoding algorithm

FER
can speed up the convergence rate. However, it degrades the −3
10
error performance in a large enough number of iterations
due to its greediness. The NWRBP resolves this problem −4
10
through generating all the outgoing messages of a check-
node simultaneously [9]. In order to choose which message −5
10
1 1.5 2 2.5 2.75
to be updated, both the RBP and the NWRBP need to Eb/No(dB)
compute several messages that are not really propagated. Such
computation greatly increases the decoding complexity in each Fig. 1. FER performance comparison of Flooding, LBP, EDS-LBP, NWRBP
iteration. In this letter, we proposed an efficient dynamic vs. 𝐸𝑏 /𝑁𝑜 for the 802.16e LDPC code with rate 1/2 and block-length 576
at maximum iteration 30.
schedule, which can greatly reduce the residual computational
complexity while achieving better error performance and faster
convergence rate. where 𝑟(𝑚𝑘 ) denotes relative message residual of variable
node 𝑘 before and after an update, 𝑓 (𝑚𝑘 ) denotes the log-
III. E FFICIENT DYNAMIC S CHEDULE
likelihood-ratio (LLR) of variable node 𝑘 after an update and
Generally speaking, dynamic decoding schedules include 𝑚𝑘 denotes the LLR of variable node 𝑘 before an update. Then
two processes: firstly choosing check nodes or variable nodes EDS-LBP chooses check nodes that connect to the variable
according to the current message state and subsequently node with the maximum 𝑟(𝑚𝑘 ) in each dynamic decoding
running a BP decoding algorithm for these check nodes or iteration. For regular LDPC codes, the number of check nodes
variable nodes. For the EDS-LBP algorithm, the first process updated in each dynamic decoding iteration is the same while
is to choose check nodes and the second is to run the LBP for irregular LDPC codes, it is different. The complete EDS-
decoding algorithm for them. In order to describe clearly, LBP algorithm is described as following:
we define these two consecutive processes as a full dynamic
1: Initialize all 𝑚𝑐→𝑣 = 0
decoding iteration. The EDS-LBP algorithm runs the dynamic
2: Initialize all 𝑚𝑣𝑗 →𝑐𝑖 = 𝐶𝑣𝑗
decoding iteration one by one until stopping rules are satis-
3: Compute all 𝑟(𝑚𝑗 )
fied. The strategy of choosing check nodes during dynamic
4: Find 𝑖 = arg max(𝑟(𝑚𝑗 ))
decoding iteration directly affects the convergence rate, error 𝑗={1,2⋅⋅⋅𝑁 }
performance and extra computational complexity. 5: For every 𝑐𝑘 ∈ 𝑁 (𝑣𝑖 ) do
In order to find an efficient strategy of choosing check 6: For every 𝑣𝑗 ∈ 𝑁 (𝑐𝑘 ) do
nodes, we exploit the following phenomena during the BP 7: Generate and propagate
decoding process: 8: Compute 𝑟(𝑚𝑗 )
1) For BP decoding process, many variable nodes achieve 9: End for
high confidence after very flew iterations. These nodes are 10: End for
not likely to change their signs or polarities in the conse- 11: Set 𝑟(𝑚𝑖 ) = 0
quent decoding iterations. The lazy schedule exploits this 12: If stopping rule is not satisfied then
phenomenon and partially updates the variable nodes based 13: Go back to position 4
on the reliability of the variable nodes message to reduce 14: End if
decoding complexity [8]. When compared to the NWRBP algorithm, the complexity
2) The differences between the messages before and after an
for computating the residual in our proposed EDS-LBP algo-
update become zero as the BP decoding converges. A message
rithm is much lower. The EDS-LBP can combine the excellent
that has a large residual means that it has not converged yet.
characteristics of the lazy schedule and the NWRBP algorithm
Therefore, propagating that message first would speed up the
together. Especially, it can exploit the RBP schedule’s excel-
decoding process. The IDS algorithm [9] is the case that it
lence in conquering trapping sets at medium to high SNR.
exploits this characteristic and speeds up the convergence rate.
Therefore, the proposed schedule can exhibit excellent error
3) According to (2), the reliability of messages from the
performance at medium to high SNR, which is shown in the
check nodes to the variable nodes is mainly dependent on
next section.
the variable node message with less reliability. Therefore, for
the messages with the same residual, propagating the message
with less reliability first would speed up the convergence rate. IV. S IMULATION R ESULTS
Based on these phenomena, the check nodes choosing To compare the performance among the EDS-LBP, the
strategy associated with the new schedule can be described Flooding, the standard LBP and the NWRBP, simulations are
as following. Firstly, we compute the residual performed on rate 1/2, length 576 and 1152 LDPC codes
 𝑓 (𝑚 ) − 𝑚  based on the IEEE 802.16e standard [11] over BIAWGN
 𝑘 𝑘
𝑟(𝑚𝑘 ) =   (4)
𝑓 (𝑚𝑘 ) + 𝑚𝑘 channel using BPSK modulation. Fig.1 shows the frame-
952 IEEE COMMUNICATIONS LETTERS, VOL. 13, NO. 12, DECEMBER 2009

0
10 maximum number of iterations 5. It is shown that the EDS-
Flooding
LBP
LBP exhibits excellent error performance at medium to high
−1
10 EDS−LBP SNR at a very few number of iterations.
NWRBP

−2
10 V. C ONCLUSION
FER

An efficient dynamic schedule for LBP decoding of LDPC


−3
10 codes is presented. It chooses the check nodes that connect
to the variable node with the maximum relative message
−4
10 residual in dynamic decoding iteration. Since this suggested
schedule combines the excellent characteristics of the lazy
−5
10
0 10 20 30 40 50 schedule and the NWRBP algorithm together, it speeds up
Number of iterations
the convergence rate and achieves better error performance
Fig. 2. FER performance comparison of Flooding, LBP and EDS-LBP,
as compared with the standard LBP algorithm. The extra
NWRBP vs. number of iterations for the 802.16e LDPC code with rate 1/2 computational complexity keeps much lower as compared with
and block-length 576 at 𝐸𝑏 /𝑁𝑜 =2.75𝑑𝐵. the NWRBP algorithm. Simulation results show that the EDS-
LBP exhibits excellent error performance at a medium to high
0
10 SNR.

−1
10 ACKNOWLEDGMENT
−2
The authors would like to thank the anonymous reviewers
10
for their helpful comments. The work was supported by the
FER

−3
National Natural Science Foundation of China (Grant No.
10
60673086, 60970041).
Flooding
−4
10 LBP
EDS−LBP R EFERENCES
NWRBP
−5
10
[1] F. R. Kschischang and B. J. Frey, “Iterative decoding of compound
2 2.1 2.2 2.3 2.4 2.5 codes by probability propagation in graphical models,” IEEE J. Sel.
Eb/No(dB) Areas Commun., vol. 16, no. 2, pp. 219-230, Feb. 1998.
[2] M. M. Mansour and N. R. Shanbhag, “High-throughput LDPC de-
Fig. 3. FER performance comparison of Flooding, LBP, EDS-LBP, NWRBP coders,” IEEE Trans. VLSI Systems., vol. 11, no. 6, pp. 976-996, Dec.
vs. 𝐸𝑏 /𝑁𝑜 for the 802.16e LDPC code with rate 1/2 and block-length 1152 2003.
at maximum iteration 5. [3] D. Hocevar, “A reduced complexity decoder architechture via layered
decoding of LDPC codes,” in Proc. IEEE Workshop on Signal Process-
ing Systems, Oct. 2004, pp. 107-112.
[4] E. Sharon, S. Litsyn, and J. Goldberger, “An efficient message-passing
error-rate (FER) performance of the different BP implemen- schedule for LDPC decoding,” in Proc. 23rd IEEE Convention of
tations for block-length 576 LDPC codes with a maximum Electrical and Electronics Engineers, Tel-Aviv, Israel, Sep. 2004, pp.
number of iterations 30. As compared with standard LBP, 223-226.
[5] H. Kfir and I. Kanter, “Parallel versus sequential updating for belief
it can be seen that the EDS-LBP greatly improves the error propagation decoding,” Physica A, vol. 330, pp. 259-270, Sep. 2003.
performance at medium to high SNR. This can be explained [6] J. Zhang and M. Fossorier, “Shuffled belief propagation decoding,” IEEE
with the fact [9] that as SNR increases, the trapping sets Trans. Commun., vol. 2, no. 53, pp. 209-213, Feb. 2005.
[7] E. Sharon, S. Litsyn, and J. Goldberger, “Efficient serial message-
errors become dominant, and RBP can delay and conquer passing schedule for LDPC decoding,” IEEE Trans. Inf. Theory, vol.
trapping sets due to its greedy iteration decoding. Fig.2 shows 53, no. 11, pp. 4076-4091, Nov. 2007.
the FER performance of different BP implementations for [8] D. Levin, E. Sharon, and S. Litsyn, “Lazy scheduling for LDPC
decoding,” IEEE Commun. Lett., vol. 11, no. 1, pp. 70-72, Jan. 2007.
block-length 576 LDPC codes as the number of iterations [9] A. Casado, M. Griot, and R. D. Wesel, “Informed dynamic scheduling
increases at 𝐸𝑏 /𝑁𝑜 = 2.75𝑑𝐵. It is shown that the EDS-LBP for belief-propagation decoding of LDPC codes,” in Proc. IEEE Int.
outperforms the standard LBP and the NWRBP in terms of Conf. on Commun, Glasgow, Scotland, June 2007, pp. 932-937.
[10] A. Casado, M. Griot, and R. D. Wesel, “Improving LDPC decoders
the convergence rate and error performance at high SNR. Fig.3 via informed dynamic scheduling,” in Proc. IEEE Information Theory
shows the frame-error-rate (FER) performance of different BP Workshop, Lake Tahoe, CA, USA, Sep. 2007, pp. 208-213.
implementations for block-length 1152 LDPC codes with a [11] LDPC coding for OFDMA PHY, IEEE C802.16e-05/066r3.

You might also like