Accelerating Minimap2 for Accurate Long Read Alignment on GPUs.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Long read sequencing is crucial for precision medicine, but current tools like Minimap2 are slow on CPUs.
• The chaining step in Minimap2 accounts for 60-70% of CPU runtime.
• Point-of-care sequencing workflows often utilize Graphics Processing Units (GPUs).

Purpose of the Study:

• To accelerate the Minimap2 sequence mapping and alignment process.
• To develop a heterogeneous computing design (mm2-ax) utilizing GPUs for the chaining step.
• To demonstrate the time and cost efficiencies of the GPU-accelerated approach.

Main Methods:

• Developed minimap2-accelerated (mm2-ax), a heterogeneous design for sequence mapping.
• Transformed Minimap2's chaining algorithm to enhance intra-read parallelism on GPUs.
• Optimized GPU utilization through high memory access, workload balancing, data locality, and reduced branch divergence.

Main Results:

• Achieved 5.41 - 2.57X speedup in the chaining step using mm2-ax on an NVIDIA A100 GPU.
• Demonstrated 4.07 - 1.93X speedup relative to cost (costup) compared to the fastest CPU version (mm2-fast).
• Maintained mapping accuracy while improving computational efficiency.

Conclusions:

• mm2-ax significantly accelerates long read sequence mapping and alignment.
• The GPU-accelerated approach offers substantial time and cost savings for precision medicine.
• Heterogeneous computing designs are effective for optimizing bioinformatics workflows on modern hardware.