Statistical Models for Automatic Performance Tuning
 Richard Vuduc,
 James W. Demmel,
 Jeff Bilmes
 … show all 3 hide
Abstract
Achieving peak performance from library subroutines usually requires extensive, machinedependent tuning by hand. Automatic tuning systems have emerged in response, and they typically operate, at compiletime, by (1) generating a large number of possible implementations of a subroutine, and (2) selecting a fast implementation by an exhaustive, empirical search. This paper applies statistical techniques to exploit the large amount of performance data collected during the search. First, we develop a heuristic for stopping an exhaustive compiletime search early if a nearoptimal implementation is found. Second, we show how to construct runtime decision rules, based on runtime inputs, for selecting from among a subset of the best implementations. We apply our methods to actual performance data collected by the PHiPAC tuning system for matrix multiply on a variety of hardware platforms.
 Title
 Statistical Models for Automatic Performance Tuning
 Book Title
 Computational Science — ICCS 2001
 Book Subtitle
 International Conference San Francisco, CA, USA, May 28–30, 2001 Proceedings, Part I
 Pages
 pp 117126
 Copyright
 2001
 DOI
 10.1007/3540455450_21
 Print ISBN
 9783540422327
 Online ISBN
 9783540455455
 Series Title
 Lecture Notes in Computer Science
 Series Volume
 2073
 Series ISSN
 03029743
 Publisher
 Springer Berlin Heidelberg
 Copyright Holder
 SpringerVerlag Berlin Heidelberg
 Additional Links
 Topics
 Industry Sectors
 eBook Packages
 Editors

 Vassil N. Alexandrov ^{(1)}
 Jack J. Dongarra ^{(2)}
 Benjoe A. Juliano ^{(3)}
 René S. Renner ^{(3)}
 C. J. Kenneth Tan ^{(4)}
 Editor Affiliations

 1. School of Computer Science, Cybernetics and Electronic Engineering, University of Reading
 2. Innovative Computing Lab, Computer Science Department, University of Tennessee
 3. Computer Science Department, California State University
 4. School of Computer Science, The Queen’s University of Belfast
 Authors

 Richard Vuduc ^{(5)}
 James W. Demmel ^{(6)}
 Jeff Bilmes ^{(7)}
 Author Affiliations

 5. Computer Science Division, University of California at Berkeley, Berkeley, CA, 94720, USA
 6. Computer Science Division and Dept. of Mathematics, University of California at Berkeley, Berkeley, CA, 94720, USA
 7. Dept. of Electrical Engineering, University of Washington, Seattle, WA, USA
Continue reading...
To view the rest of this content please follow the download PDF link above.