Our Pangenome Layout Paper is Accepted by SC24.
In our paper, we address the challenge of visualizing large pangenome graphs, which are essential for understanding genetic diversity. We analyze a state-of-the-art layout algorithm and uncover significant data-level parallelism, highlighting the potential of GPU acceleration. Despite facing challenges like irregular data access and memory-bound behavior, we develop an optimized GPU-based solution using a cache-friendly data layout, coalesced random states, and warp merging. We also introduce a new metric for evaluating layout quality at scale. Our approach achieves a 57.3x speedup over the best multithreaded CPU baseline on 24 human whole-chromosome pangenomes, reducing execution time from hours to minutes without compromising layout quality.
Yixiao Du
publication