Scientific Computing Seminar

Date:

Thursday, January 26, 2006

Time:

1:00pm-2:00pm

Location:

70-191

Seminar Speaker:

Alan Snavely
San Diego Supercomputer Center
http://www.sdsc.edu/~allans/allans.html

Title:

Methods for Solving the Convolution Problem in Performance Modeling

Abstract:

Previous work has described the convolution problem in performance modeling as follows: machine profiles that are rates at which a computer can perform operations in various categories are combined with application signatures that are counts of the operations performed by an application in each category to yield a performance prediction.
In this work we more formally define performance convolutions as mathematical objects, then present several methods for solving them to obtain performance predictions. Each of these methods has practical advantages and disadvantages with respect to difficulty in evaluating the convolution, subsequent accuracy of predictions, and depending on the availability of the input data required to solve the convolution.
Also explored are methods for "ranking supercomputer rankings" based on simple metrics to determine how well they correspond to real application's performance.

Sponsor of Seminar:

Erich Strohmaier

Scientific Computing

Contact Esmond G. Ng EGNg@lbl.gov