Matúš Liščinský: Performance Analysis Based on Noise Injection

The goal of this thesis was to extend the capabilities of Perun with injecting noise into program to either simulate potential code optimisations or to make performance issues manifest during profiling.

Abstract

In this work, we proposed a Perun-Blower framework which utilises the perfblowing technique: injecting of noise into the functions of the tested program, followed by collecting of runtime data of these functions from the program run and evaluating the impact of the noise on the program performance. We build on the dynamic binary instrumentation of the Pin framework to inject the noise into program. We then focus on finding functions with high impact on performance as well as estimate the thread run’s potential acceleration when optimising the particular functions. Moreover, we have extended the existing Trace collector used in the Perun framework to collect the runtime of functions with a new so-called engine based on the Pin framework. We tested the functionality of our implementation on two non-trivial projects, where we were able to find functions (1) with considerable impact on performance, (2) with the most significant optimisation benefit, and (3) whose degradation forces the non-termination of the program after several hours of running.