Statistical Models for Automatic Performance Tuning

Achieving peak performance from library subroutines usually requires extensive, machine-dependent tuning by hand. Automatic tuning systems have emerged in response, and they typically operate by (1) generating a large number of possible implementations of a subroutine, and (2) selecting the fastest implementation by an exhaustive, empirical search. This paper presents quantitative data that motivates the development of such a search-based system, and discusses two problems which arise in the context of search. First, we develop a heuristic for stopping an exhaustive compile-time search early if a near-optimal implementation is found. Second, we show how to construct run-time decision rules, based on run-time inputs, for selecting from among a subset of the best implementations. We address both problems by using statistical techniques to exploit the large amount of performance data collected during the search. We apply our methods to actual performance data collected by the PHiPAC tuni...

• Type: Conference
• Year: 2001
• Where: ICCS
• Authors: Rich Vuduc, James Demmel, Jeff Bilmes
• Keywords: Applied Computing, Automatic Tuning Systems, Exhaustive Compile-time Search, ICCS 2001, Machine-dependent Tuning