GPU-REMuSiC on Graphics Processing Units

Chun-Yuan Lin, Assistant Processor

Department of Computer Science and Information Engineering

Chang Gung University

GPU-REMuSiC on Graphics

Processing Units

2011/5/20 1 PPCB

Outline

Sequence alignment problem

Pair-wise sequence alignment

Multiple sequence alignment

Previous studies

Parallel computing for dynamic programming

Related works on CUDA

Constrain multiple sequence alignment

Developments of a series of tools

RE-MuSiC

GPU-REMuSiC implementation

2011/5/20 2 PPCB

Sequence alignment problem

2011/5/20 3 PPCB

2011/5/20 4 PPCB

2011/5/20 5 PPCB

2011/5/20 6 PPCB

2011/5/20 7 PPCB

2011/5/20 8 PPCB

2011/5/20 9 PPCB

2011/5/20 10 PPCB

PAM100

2011/5/20 11 PPCB

2011/5/20 12 PPCB

2011/5/20 13 PPCB

2011/5/20 14 PPCB

2011/5/20 15 PPCB

2011/5/20 16 PPCB

2011/5/20 17 PPCB

We can not use dynamic programming for k sequences.

2011/5/20 18 PPCB

2011/5/20 19 PPCB

2011/5/20 20 PPCB

Previous studies

2011/5/20 21 PPCB

2011/5/20 22 PPCB

2011/5/20 23 PPCB

Some bioinformatics applications have been successfully

ported to GPGPU in the past.

Liu et al. (IPDPS 2006) implemented the Smith-Waterman algorithm to

run on the nVidia GeForce 6800 GTO and GeForce 7800 GTX, and

reported an approximate 16× speedup by computing the alignment

score of multiple cells simultaneously.

Charalambous et al. (LNCS 2005) ported an expensive loop from

RAxML, an application for phylogenetic tree construction, and achieved a

1.2× speedup on the nVidia GeForce 5700 LE.

2011/5/20 24 PPCB

Liu et al. (IEEE TPDS 2007) presented a GPGPU approach to

high-performance biological sequence alignment based on

commodity PC graphics hardware. (C++ and OpenGL Shading

Language (GLSL))

Pairwise Sequence Alignment (Smith-Waterman algorithm, scan database)

Multiple sequence alignment (MSA)

2011/5/20 25 PPCB

Some bioinformatics applications have been successfully

ported to CUDA now.

Smith-Waterman algorithm (goal: scan database)

Manavski and Valle (BMC Bioinformatics 2008),

Striemer and Akoglu (IPDPS 2009),

Liu et al. (BMC Research Notes 2009)

Khajeh-Saeed et al. (Journal of Computational Physics, 2010)

Liu et al. (BMC Research Notes 2010)

Multiple sequence alignment (ClustalW)

Liu et al. (IPDPS 2009) for Neighbor-Joining Trees construction

Liu et al. (ASAP 2009)

2011/5/20 26 PPCB

Manavski and Valle present the first solution based on commodity hardware that efficiently computes the exact Smith-Waterman alignment. It runs from 2 to 30 times faster than any previous implementation on general-purpose hardware.

Pre-compute a query profile parallel to the query sequence for each possible residue.

The implementation in CUDA was to make each GPU thread compute the whole alignment of the query sequence with one database sequence. (pre-order the sequences of the database in function of their length)

The ordered database is stored in the global memory, while the query-profile is saved into the texture memory.

For each alignment the matrix is computed column by column in order parallel to the query sequence. (store them in the local memory of the thread)

2011/5/20 27 PPCB

Striemer and Akoglu further study the effect of memory organization and

the instruction set architecture on GPU performance.

They pointed out that query profile in Manavski’s method has a major drawback

in utilizing the texture memory of the GPU that leads to unnecessary caches

misses. (larger than 8KB)

Long sequence problem.

They placed the substitution matrix in the constant memory to exploit the

constant cache, and created an efficient cost function to access it. (modulo

operator is extremely inefficient on CUDA, not use hash function)

They mapped query sequence as well as the substitution matrix to the constant

memory.

2011/5/20 28 PPCB

Liu et al. proposed Two versions of CUDASW++ are implemented: a

single-GPU version and a multi-GPU version.

The alignment can be computed in minor-diagonal order from the top-left

corner to the bottom-right corner in the alignment matrix.

Considering the optimal local alignment of a query sequence and a subject

sequence as a task.

Inter-task parallelization: Each task is assigned to exactly one thread and dimBlock

tasks are performed in parallel by different threads in a thread block.

Intra-task parallelization: Each task is assigned to one thread block and all dimBlock

threads in the thread block cooperate to perform the task in parallel.

2011/5/20 29 PPCB

Khajeh-Saeed et al. showed that effective use of the GPU requires a novel

reformulation of the Smith–Waterman algorithm.

2011/5/20 30 PPCB

Liu et al. described the latest release of the CUDASW++ software,

CUDASW++ 2.0, which makes new contributions to Smith-Waterman

protein database searches using compute unified device architecture

(CUDA).

2011/5/20 31 PPCB

Liu et al. presents MSA-CUDA, a parallel MSA program, which

parallelizes all three stages of the ClustalW processing pipeline using

CUDA.

Pairwise distance computation:

As the work in Liu et al. (BMC Research Notes 2009)

Neighbor-Joining Trees: as the work in Liu et al. (IPDPS 2009)

Progressive alignment:

conducted iteratively in a

multi-pass way.

2011/5/20 32 PPCB

Constrain multiple sequence alignment

2011/5/20 33 PPCB

(Tang, CSB 2002) P = HKH

Constrain Multiple Sequence Alignment (CMSA)

2011/5/20 34 PPCB

G. Myers, S. Selznick, Z. Zhang and W. Miller. “Progressive multiple alignment with constraints,” Proceedings of the 1st ACM Conference on Computational Molecular Biology, 1997.

(Position constraint, is jt)

C.Y. Tang, C.L. Lu, M.D.T. Chang, Y.T. Tsai, etc, “Constrained Multiple Sequence Alignment Tool Development and Its Application to RNase Family Alignment,” Journal of Bioinformatics and Computational Biology, vol. 1, 2003. (CSB 2002) (CMSA)

(Character constraint, P: a set of single residue/nucleotide)

Y.T. Tsai, Y.P. Huang, C.T. Yu and C.L. Lu, “MuSiC: a tool for multiple sequence alignment with constraints,” Bioinformatics, vol. 20, 2004. (MuSiC)

(Segment constraint, P: a set of short segment) C.L. Lu and Y.P. Huang, “A memory-efficient algorithm for multiple

sequence alignment with constraints,” Bioinformatics, vol. 21, 2005. (MuSiC-ME)

2011/5/20 35 PPCB

Y.-T. Tsai, “The constrained common sequence problem,” Information Processing Letters, vol. 88, 2003.

F. Y.L. Chin, etc,

“A simple algorithm for the constrained sequence problems,” IPL, vol. 90, 2004.

“ Efficient Constrained Multiple Sequence Alignment with Performance Guarantee,” JBCB, vol. 3, 2005. (CSB, 2003)

A. N. Arslan, etc.

“Algorithms for the constrained common sequence problem,” Proc. Prague Stringology Conference, 2004.

“A fast Algorithm for the Constrained Multiple Sequence Alignment problem,” Acta Cybernetica (accepted)

“A parallel algorithm for the constrained multiple sequence alignment problem,” BIBE 2005.

“A space-efficient algorithm for the constrained pairwise sequence alignment problem,” GIW 2005.

“FastPCMSA: An improved parallel algorithm for the constrained multiple sequence alignment problem,” FCS 2006.

“A* Algorithms for the Constrained Multiple Sequence Alignment Problem,” ICAI 2006.

“Space-efficient Parallel Algorithms for the Constrained Multiple Sequence Alignment Problem,” BIOCOMP 2006.

2011/5/20 36 PPCB

2011/5/20 37 PPCB

RE-MuSiC steps Step1. Compute the score of the global sequence alignment without

any constraint using the Needleman-Wunsch algorithm between all pairs of the K sequences and build the distant matrix.

Step2. Follow the results of Step1 to create a complete graph G of K sequences.

Step3. Use the complete graph G to construct a Kruskal merging order tree .

Step4. Use the created in Step3 as a guide tree. Progressively align the sequences according to the branching order of the Two closest pre-aligned groups of sequences are joined by CPSA algorithm to represented sequences of two groups. (ClustalW)

2011/5/20 38 PPCB

GPU-REMuSiC implementation

2011/5/20 39 PPCB

In the past, we have designed an efficient parallel algorithm for the

constrained multiple sequence alignment based on the memory-efficient

algorithm (MuSiC-ME). (PCMSA, MPI)

3k

0

5

10

15

20

25

30

0 5 10 15 20 25 30 35

total processers

spee

d up

32 sequences 64 sequences 128 sequences

256 sequences 512 sequences

4k

0

5

10

15

20

25

30

0 5 10 15 20 25 30 35

total processers

spee

d up



5k

0

5

10

15

20

25

30

0 5 10 15 20 25 30 35

total processers

spee

d u

p



2011/5/20 40 PPCB

Define eight possible

methods for computing

dynamic programming

by CUDA

(a)SRST (b) SRMT

Figure 3: SRST and SRMT.

(a) ARST (b) ARMT

Figure 4: ARST and ARMT.

(a) SDST (b) SDMT

Figure 5: SDST and SDMT.

(a) ADST (b) ADMT

Figure 6: ADST and ADMT.

2011/5/20 41 PPCB

Figure 7. Distance matrix and the thread blocks allocation.

Figure 8. The load balancing cutting technology by two GPUs

2011/5/20 42 PPCB

(Device) Grid

Constant

Memory

Texture

Memory

Global

Memory

Block (0, 0)

Shared Memory

Local

Memory

Thread (0, 0)

Registers

Local

Memory

Thread (1, 0)

Registers

Block (1, 0)

Shared Memory

Local

Memory

Thread (0, 0)

Registers

Local

Memory

Thread (1, 0)

Registers

Host

Intra-task parallelization

•Sequences (global memory)

•Scoring matrix (constant memory)

•alphabetical order

•ASCII indexing

•Si and Sj stored in shared memory

•Result stored in shared memory

•Memory efficient

•Only score

2011/5/20 43 PPCB

Sp

eedu

ps

Sequences

90.4 85.75 92.39 87.63

179.24 177.27 187.38 181.17

229.58

268.1 255.41273.1272.04

374.86342.78

383.14

0

50

100

150

200

250

300

350

400

450

400(856) 2000(266) 1000(858) 2000(858)

1 GPU 2 GPU 3 GPU 4 GPU

(DP computation time)

2011/5/20 44 PPCB

Sp

eedu

ps

Sequences

Figure 10. The speedup ratios by comparing GPU-REMuSiC and RE-MuSiC.

12.51

22.78

12.59 13.5113.3

26.36

13.5

21.1

13.5

28.05

13.68

21.9

13.52

29.18

13.93

22.36

0

5

10

15

20

25

30

35

400(856) 2000(266) 1000(858) 2000(858)


2011/5/20 45 PPCB

0%

10%

20%

30%

40%

50%

60%

70%

80%

90%

100%

CPU 1 GPU 2 GPU 3 GPU 4 GPU

Other Distant Matrix

2011/5/20 46 PPCB

Sp

eedu

ps

Length

0

50

100

150

200

250

300

350

0 100 200 300 400 500 600 700


2011/5/20 47 PPCB

GPU-REMuSiC on Graphics Processing Units

Documents