Performance Variance Evaluation on Mozilla Firefox

Abstract

In order to evaluate software performance and find regressions, many developers use automated performance tests. However, the test results often contain a certain amount of noise that is not caused by actual performance changes in the programs. They are instead caused by external factors like operating system decisions or unexpected non-determinisms inside the programs. This makes interpreting the test results hard since results that differ from previous results cannot easily be attributed to either genuine changes or noise. In this thesis we use Mozilla Firefox as an example to try to find the causes for this performance variance, develop ways to reduce the noise and present a statistical technique that makes identifying genuine performance changes more reliable. Our results show that a significant amount of noise is caused by memory randomization and other external factors, that there is variance in Firefox internals that does not seem to be correlated with test result variance, and that our suggested statistical forecasting technique can give more reliable detection of genuine performance changes than the one currently in use by Mozilla.