RabbitSAlign: Accelerating Short-Read Alignment for CPU-GPU Heterogeneous Platforms

Short-read alignment is a critical, yet time-consuming step in many next-generation sequencing data analysis pipelines. Most approaches follow the seed-and-extend strategy, where seeding usually involves a large number of random memory accesses, and extension of seeds relies on computationally expensive alignment algorithms, resulting in huge time consumption. Recently, Strobealign has reached state-of-the-art alignment speed while maintaining high accuracy through an innovative seeding strategy. Yet, there is still room for further optimization, especially on modern CPU-GPU heterogeneous platforms. In this paper, we present RabbitSAlign, a new GPU-accelerated short-read aligner based on Strobealign. By optimizing inefficient operations in the seeding process and utilizing GPUs to accelerate the extension process, RabbitSAlign doubles the processing speed on real biological datasets compared to Strobealign. It surpasses the performance of highly optimized BWA-MEM2 and NVIDIA Parabricks by a factor of at least four, while also being one-order-of-magnitude faster than the widely-utilized BWA-MEM and Bowtie2. Additionally, RabbitSAlign features highly competitive accuracy on both simulated and real biological data. Remarkably, it can process a 30x human genome sequencing dataset in merely 18 min. C++ sources are available at https://github.com/RabbitBio/RabbitSAlign.