University of Minnesota
University Relations
http://www.umn.edu/urelate
612-624-6868
myU OneStop


Go to unit's home.

Home | Seminars and Symposia | Past seminars/symposia: Wednesday, November 12, 2003

DTC Seminar Series

Performance Metrics Based on Computational Least Action

by

Robert W. Numrich
Senior Research Associate
University of Minnesota
Supercomputing Institute

Wednesday, November 12, 2003
1:00 pm

402 Walter Library

Robert W. NumrichRenewed interest in computational metrics has been stimulated by the High Productivity Computing Systems (HPCS) project sponsored by DARPA. Substitution of the word "productivity" for the more common word "performance" is a deliberate statement from DARPA that software development time is just as important, and always has been just as important, as program execution time. The HPCS project is looking for new metrics that measure both aspects of productivity with the goal of minimizing the "total-time-to-solution" for scientific computing. As an approach to this problem, Dr. Numrich will first discuss what it might mean to minimize total-time-to-solution. Secondly, he will turn the focus of attention away from time alone as a metric to look more at the work performed during a project within the time constraints imposed on the project. This puts the project into the energy-time plane whose area has dimensions of what Dr. Numrich calls computational action. Dr. Numrich suggests that the Principle of Computational Least Action is a new framework for optimizing total-time-to-solution. He will illustrate this idea with a simple example using the classical methods of the calculus of variations.

Numrich presenting

 

Bob Numrich is a Senior Research Associate at the Minnesota Supercomputing Institute. He has been involved in high-performance computing for many years starting as a graduate student at the University of Minnesota working with Don Truhlar. He was one of the pioneers of vector computing first at Control Data and then at Cray Research. He was a member of the design team for Cray's massively parallel machine, the CRAY-T3D/E, for which he invented the one-sided programming model, which is now known as the Shmem model. He also designed a simple parallel extension to Fortran 90, called Co-Array Fortran, which incorporates an enhanced version of the Shmem model into the language. His interest in performance modeling began with simulations of memory contention for the Cray-2 computer. The principle of computational least action follows directly from an analogy with classical mechanics first used as a model to explain those simulations.