Database peptide search is the primary computational technique to deduce peptides from the mass-spectrometry data. However, even with significant algorithmic and computational advances in database peptide search methods, the existing algorithms exhibit sub-optimal performance and limited GPU application. This mainly stems from their inefficient parallel designs and computational methods, and high overhead costs.
We present HiCOPS and its GPU-accelerated version, GiCOPS, for scalable acceleration of database peptide search on modern supercomputers. The HiCOPS framework presents a pipeline of parallel algorithms, data structures, and optimizations to optimally distribute and compute the database search across parallel nodes. GiCOPS further extends and accelerates the HiCOPS's core computational algorithms using GPUs. HiCOPS outperforms several existing algorithms by more than 10x (strong-scale efficiency: 70-80%). GiCOPS further speeds up the search by 4-5x. We assess the performance of both frameworks and report near-optimal results for several metrics including load-balance, compute and memory throughputs, and performance overheads.