CUDASW++ 3.0: accelerating Smith-Waterman protein database search by coupling CPU and GPU SIMD instructions.

Area of Science:

• Bioinformatics
• Computational Biology
• High-Performance Computing

Background:

• The Smith-Waterman algorithm is crucial for sensitive protein sequence database searches.
• Its quadratic time complexity and growing databases present significant computational challenges.

Purpose of the Study:

• To develop a faster Smith-Waterman algorithm for protein sequence database searching.
• To leverage combined CPU and GPU resources for enhanced computational performance.

Main Methods:

• Implemented CUDASW++ 3.0 utilizing CPU SSE-based vector units and GPU CUDA PTX SIMD instructions.
• Introduced concurrent CPU and GPU computations and automatic workload distribution.
• Developed a novel GPU SIMD parallelization beyond the standard SIMT model.

Main Results:

• CUDASW++ 3.0 achieved performance improvements of up to 2.9x and 3.2x over CUDASW++ 2.0.
• Reached maximum performance rates of 119.0 and 185.6 GCUPS on single and dual GPUs, respectively.
• Demonstrated significant speedups compared to SWIPE and BLAST+.

Conclusions:

• CUDASW++ 3.0 offers substantial speedups by integrating CPU/GPU SIMD instructions and concurrent execution.
• The algorithm is optimized for Kepler architecture GPUs and written in CUDA C++/PTX assembly.
• Source code and data are publicly available for further research and application.