PNNL

Abstract

We have conducted a detailed study to understand the potential of hybrid CPU/FPGA High-Performance Computers for improving the performance of scientific applications. In particular, we have focused on an application in proteomics (Polygraph ), which is representative of many types of computational analysis applications in the life sciences: it focuses on extracting useful information from a large body of experimentally collected data (identifying observed peptide spectra collected from a mass spectrometer against a well-known protein database). Our preliminary analysis of Polygraph found that more than half (> 51%) of the computation was spent in three routines. We have implemented an FPGA version of the most computationally-intensive routine on a Cray XD-1 system, and measured the overall speedup achieved in comparison to an optimized software version of the routine running on the Cray XD-1’s native Opteron processors. We have achieved computational speedups of up to 9.16. When we include data movement costs, the overall speedups is reduced to 1.78. We discuss the design and implementation strategies that led to these results, as well as advantages and limitations we found on the Cray XD-1 platform. We also address the advantages and limitations of current development environments, as well as discuss relevant issues we found in our experience as hybrid CPU/FPGA programming model “users”.