High-performance computing using GPUs: examples from computational biology

Bharat Sukhwani

PhD candidate, Electrical and Computer Engineering Department

Boston University

Abstract

1. Graphics processing units have long offered very high raw floating point capabilities. Until recently, however, their application to high performance computing had been limited due to their relatively fixed architecture and the difficulties in mapping general computing to graphics processors. With the advent of more flexible GPU architectures and high-level GPU programming languages, GPUs have seen wide acceptance in the HPC community. Modern GPUs offer multi GFLOPs to a TFLOP of peak performance at a fraction of the cost of the large clusters.
   

2. Although programming the GPUs has become relatively easier, achieving high performance from them still requires a good understanding of the underlying hardware architecture and the memory hierarchy as well as efficient mapping of the computations to the GPU threads.
   

3. In this talk, we discuss our experience of mapping two computational biology applications on graphics processors - molecular docking and binding-site mapping. These applications are computationally very demanding, requiring many hours to days of runtime on conventional CPUs. We show how these computations can be mapped to the GPUs and benefit from the high performance offered by them.
   

About the speaker

1. Bharat Sukhwani did his Bachelor of Engineering in Bombay University, and his M.S. in University of Arizona, Tucson.  He currently works under Professor Martin Herbordt in the field of parallel computing, use of application-specific and reconfigurable hardware for HPC, and electronic design automation.