News

Seminar

The Relevance of New Data Structure Approaches for Dense Linear Algebra in the new Multi-Core / Many Core Environments

Fred G. Gustavson
IBM T.J. Watson Research Center

March 4, 2008
Location: Lally 102
Time: 4:00 pm to 5:00 pm

Abstract:

For about ten years now, Bo Kagstroms Group in Umea, Sweden, Jerzy Wasniewskis team at Danish Technical University in Lyngby, Denmark, John Gunnels and I at IBM Research in Yorktown Heights have been applying recursion and New Data Structures (NDS) to increase the performance of Dense Linear Algebra (DLA) factorization algorithms (FA). For about three years now almost all computer manufacturers have dramatically changed their computer architectures which they call Multi-Core, (MC). We demonstrate that many DLAFA can be viewed almost entirely as performing matrix multiplication with update, DGEMM. The perfomance of current LAPACK and ScaLAPACK depend crucially on Level 3 Basic Linear Algebra Subroutines (BLAS) and PBLAS. However, these industry standards require the same Application Program Interface (API) that LAPACK and ScaLAPACK does, namely standard full and packed storage formats for dense matrices. These standard formats are the crux of poor performance for LAPACK and ScaLAPACK. This talk will explain why and how this perfomance can be improved using NDS. It turns out that the new MC designs give poor performance for the existing traditional designs of DLA libraries such as LAPACK and ScaLAPACK. Recent results of Jack Dongarras group at the Innovative Computing Laboratory in Knoxville, Tennesee have shown how to obtain high performance for DLAFA algorithms on the Cell architecture, an example of an MC processor, but only when they used NDS. Thus this new work gives an experimental verification of the subject matter of this talk.

Hosted by: Boleslaw K. Szymanski (x2714)
Administrative support: Chris Coonrad (x8412)

Last updated: January 18, 2008