Title: Communication-avoiding linear algebra

Speaker: Mark Hoemmen, UC Berkeley

Date/Time: Monday, March 16, 2009 – 9:30 AM

Location: CSRI Building, Room 90 (Sandia NM)

Brief Abstract: Moving data between processors or levels of a memory hierarchy is often orders of magnitude more time-consuming than arithmetic operations, and this disparity has gotten exponentially worse over time.  I will explain how important dense and sparse linear algebra algorithms have communication bottlenecks which hinder scalability at all levels, from a single multicore node to a massively parallel supercomputer.  Then, I will present new algorithms which communicate asymptotically less, without sacrificing accuracy.  Performance models and benchmark results show the practical promise of the new algorithms.

CSRI POC: Scott Collis, (505) 284-1123