CUDAMPF++: A Proactive Resource Exhaustion Scheme for Accelerating Homologous Sequence Search on CUDA-enabled GPU

摘要

Genomic sequence alignment is an important research topic in bioinformaticsand continues to attract significant efforts. As genomic data growexponentially, however, most of alignment methods face challenges due to theirhuge computational costs. HMMER, a suite of bioinformatics tools, is widelyused for the analysis of homologous protein and nucleotide sequences with highsensitivity, based on profile hidden Markov models (HMMs). Its latest version,HMMER3, introdues a heuristic pipeline to accelerate the alignment process,which is carried out on central processing units (CPUs) with the support ofstreaming SIMD extensions (SSE) instructions. Few acceleration results havesince been reported based on HMMER3. In this paper, we propose a five-tieredparallel framework, CUDAMPF++, to accelerate the most computationally intensivestages of HMMER3's pipeline, multiple/single segment Viterbi (MSV/SSV), on asingle graphics processing unit (GPU). As an architecture-aware design, theproposed framework aims to fully utilize hardware resources via exploitingfiner-grained parallelism (multi-sequence alignment) compared with itspredecessor (CUDAMPF). In addition, we propose a novel method that proactivelysacrifices L1 Cache Hit Ratio (CHR) to get improved performance and scalabilityin return. A comprehensive evaluation shows that the proposed frameworkoutperfroms all existig work and exhibits good consistency in performanceregardless of the variation of query models or protein sequence datasets. ForMSV (SSV) kernels, the peak performance of the CUDAMPF++ is 283.9 (471.7) GCUPSon a single K40 GPU, and impressive speedups ranging from 1.x (1.7x) to 168.3x(160.7x) are achieved over the CPU-based implementation (16 cores, 32 threads).