CUDASW++2.0: enhanced Smith-Waterman protein database search on CUDA-enabled GPUs based on SIMT and virtualized SIMD abstractions

Abstract

BackgroundDue to its high sensitivity, the Smith-Waterman algorithm is widely used for biological database searches. Unfortunately, the quadratic time complexity of this algorithm makes it highly time-consuming. The exponential growth of biological databases further deteriorates the situation. To accelerate this algorithm, many efforts have been made to develop techniques in high performance architectures, especially the recently emerging many-core architectures and their associated programming models.FindingsThis paper describes 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). A parallel Smith-Waterman algorithm is proposed to further optimize the performance of CUDASW++ 1.0 based on the single instruction, multiple thread (SIMT) abstraction. For the first time, we have investigated a partitioned vectorized Smith-Waterman algorithm using CUDA based on the virtualized single instruction, multiple data (SIMD) abstraction. The optimized SIMT and the partitioned vectorized algorithms were benchmarked, and remarkably, have similar performance characteristics. CUDASW++ 2.0 achieves performance improvement over CUDASW++ 1.0 as much as 1.74 (1.72) times using the optimized SIMT algorithm and up to 1.77 (1.66) times using the partitioned vectorized algorithm, with a performance of up to 17 (30) billion cells update per second (GCUPS) on a single-GPU GeForce GTX 280 (dual-GPU GeForce GTX 295) graphics card.ConclusionsCUDASW++ 2.0 is publicly available open-source software, written in CUDA and C++ programming languages. It obtains significant performance improvement over CUDASW++ 1.0 using either the optimized SIMT algorithm or the partitioned vectorized algorithm for Smith-Waterman protein database searches by fully exploiting the compute capability of commonly used CUDA-enabled low-cost GPUs.

An Overview of Hardware-Based Acceleration of Biological Sequence Alignment

“…Device Database Performance searched (Liu, Schmidt, Voss, Schroder & Muller-Wittig, 2006) GTX 7800 Swiss-Prot 650 MCUPS (Liu, Huang, Johnson & Vaidya, 2006) GTX 7800 983 protein 241 MCUPS sequences (Manavski & Valle, 2008) GTX 8800 Swiss-Prot 1.9 GCUPS (Liu et al, 2009) GTX 280 Swiss-Prot 9.5 GCUPS (Liu et al, 2010) GTX 280 Swiss-Prot 17 GCUPS (Kentie, 2010) GTX 275 Swiss-Prot 21.4 GCUPS …”

“…The 'CUDASW++' solution presented in (Liu et al, 2009) manages a maximum speed of about 9.5 GCUPS searching Swiss-Prot on a Geforce GTX 280 graphics card. An improved version, 'CUDASW++ 2.0' has been published recently (Liu et al, 2010). Being the fastest Smith-Waterman GPU implementation to date, 'CUDASW++ 2.0' managed 17 GCUPS on a single GTX 280 GPU, outperforming CPU-based BLAST in its benchmarks.…”

G-DNA – a highly efficient multi-GPU/MPI tool for aligning nucleotide reads

“…In contrast to the already available implementations, e.g. [6][7][8][9], our solution is highly optimized for reads from modern sequencers (Roche/454, Illumina/Solexa and AB/SOLiD) and, unlike cited works, directly targets the described problem. The software may prove to be useful especially in processing data from Roche/454, where the overlap-layout-consensus method is the most commonly employed.…”

“…Liu Y. et al [20], in turn, developed the CUDASW++ algorithm achieving up to 16GCUPS on a dual-GPU GeForce GTX 295. This approach was further explored by its authors resulting in an optimized SIMT and partitioned vectorized algorithm CUDASW++ 2.0 [7] with a very good performance * e-mail: michal.kierzynka@cs.put.poznan.pl of up to 17 GCUPS on a GeForce GTX 280 and 30 GCUPS on a dual-GPU GeForce GTX 295. On the CPU side, one solution that deserves special attention is SWIPE developed by Rognes [9].…”

Pairwise sequence alignment for very long sequences on GPUs

“…However, for large sequences, this cannot be easily achieved as only a portion of the computations can be stored in the SM memory. For example, our experiments indicate that CUDASW + +2.0 [8] cannot handle strings whose length is more than 70000 on NVIDIA Tesla C2050.…”