home · mobile · calendar · colloquia · 2008-2009 · 

Colloquium - Walters

HPC: From Massively Multicore Processors to Supercomputers
University of Buffalo

Scientific computing is experiencing an explosion in computational power. This has come in the form of both increasingly powerful computational clusters as well as massive performance increases at the desktop level. Computations that traditionally require days of compute time on expensive cluster infrastructure are now accessible using the massively multicore GPU in our desktop. At the same time, clusters are getting larger and more commonplace, bringing issues of stability and fault tolerance to the forefront. In this talk, I will present two aspects of my work. The first portion of my talk will focus on the design and implementation of GPU-HMMER, a popular CUDA-based implementation of the HMMER protein sequence analysis suite. Through careful optimization, we have found that GPUHMMER is capable of outperforming mid-sized clusters using up to three graphics processors and only a single computational node. The second portion of my talk will focus on improving the performance of MPI-based checkpointing by using a novel local disk with replication strategy. By storing checkpoints to local disks, we can avoid saturating a cluster's shared storage infrastructure. I show that by using a replication-based local storage solution, we can substantially reduce the overhead of MPI-based checkpointing.

This talk is sponsored by the National Center for Atmospheric Research Computational & Information Systems Laboratory and will be held in the Visualization Lab at the Mesa Lab.

Department of Computer Science
University of Colorado Boulder
Boulder, CO 80309-0430 USA
May 5, 2012 (14:13)