BowMapCL: Burrows-Wheeler Mapping on Multiple Heterogeneous Accelerators.

Area of Science:

• Bioinformatics
• Computational Biology
• High-Performance Computing

Background:

• Exact-search alignment methods face computational demands, often limiting problem size and implementation due to parallel processing architectures.
• Existing solutions may not efficiently utilize diverse hardware like multiple GPUs or APUs.

Purpose of the Study:

• To introduce BowMapCL, a novel exact-search alignment tool designed for efficient execution on heterogeneous parallel architectures.
• To overcome limitations of current tools by enabling unified implementation across different devices using OpenCL.

Main Methods:

• Developed BowMapCL utilizing the Burrows-Wheeler Transform (BWT) and FM-Index.
• Implemented using OpenCL for cross-platform compatibility on various GPUs and APUs.
• Incorporated performance and scalability techniques: multiple buffering, work-queue task distribution, dynamic load balancing, index partitioning, bit-encoding, and sampling.

Main Results:

• BowMapCL (single GPU) is 2x-7.5x faster than leading multi-threaded CPU BWT aligners (Bowtie, BWA, SOAP2).
• Achieved up to 4x speedup compared to state-of-the-art GPU aligners (SOAP3, HPG-BWT).
• Demonstrated efficient throughput scaling and load balancing across multiple distinct devices.

Conclusions:

• BowMapCL offers a significant performance improvement for exact-search sequence alignment.
• Its unified OpenCL implementation effectively exploits heterogeneous computing resources.
• The tool provides a scalable and efficient solution for large-scale bioinformatics analyses.