home · mobile · calendar · colloquia · 1996-1997 · 

Colloquium - Jessup

Building Efficient Parallel Linear Algebra Software
Department of Computer Science

Three fundamental characteristics of an efficient parallel program are low interprocessor data communication costs, an even balance of processor workload, and good performance of the serial program segments on the parallel computer's individual processors. In this talk, we outline the process of converting a linear algebra algorithm into an efficient parallel program. We use Householder tridiagonalization as a case study to illustrate the relative impacts of the three aforementioned characteristics on program performance. We show how innovative data distribution patterns can overcome performance problems associated with standard parallel linear algebra implementations.

Refreshments will be served immediately before the talk at 3:30pm.
Hosted by Dirk Grunwald.

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