16th International Conference on Parallel Architecture and Compilation Techniques (PACT 2007) (2007)
Sept. 15, 2007 to Sept. 19, 2007
DOI Bookmark: http://doi.ieeecomputersociety.org/10.1109/PACT.2007.7
Jeremy Lau , University of California, San Diego, USA
Matthew Arnold , IBM T.J. Watson Research Center, USA
Michael Hind , IBM T.J. Watson Research Center, USA
Brad Calder , University of California, San Diego, USA
Performance auditing is an online optimization strategy that empirically measures the effectiveness of an optimization on a particular code region. It has the potential to greatly improve performance and prevent degradations due to compiler optimizations. Performance auditing relies on the ability to obtain sufficiently many timings of the region of code to make statistically valid conclusions. This work extends the state-of-the-art of performance auditing systems by allowing a finer level of granularity for obtaining timings and thus, increases the overall effectiveness of a performance auditing system. The problem solved by our technique is an instance of the general problem of correlating a program?s high-level behavior with its binary instructions, and thus, can have uses beyond a performance auditing system. We present our implementation and evaluation of our technique in a production Java VM.
B. Calder, M. Hind, J. Lau and M. Arnold, "A Loop Correlation Technique to Improve Performance Auditing," 16th International Conference on Parallel Architecture and Compilation Techniques (PACT 2007)(PACT), Brasov, Romania, 2007, pp. 259-269.