On the precision of precise event based sampling

Abstract

Many performance studies rely on Intel's Precise Event Based Sampling (PEBS) to collect processor events, where precision is a key for the reliability of analysis. In this paper, we make a study on the precision of PEBS and show that, while by its name being precise, PEBS can cause mistakes under shadowing, which may make the analysis unreliable. We then show how to remedy such imprecision by artificially inserting bogus instructions. Evaluation shows that our remedy leads to more precise event samples and thus more reliable performance analysis.