Abstract

In today's software industry, the conflict between the need for high-quality profiles and the lack of time for profile collection hinders the widespread use of Profile-Based Optimization (PBO). We propose a methodology called progressive profiling, which tackles the realities of the daily build cycle by significantly reducing the time required for profiling runs. The approach is based on a combination of techniques for profile propagation and selective profile collection. In particular, we introduce a new algorithm for binary matching that not only allows us to reuse profile information from older versions of the program, but also yields a confidence value that we can use to choose between propagation and selective collection. Based on this algorithm we implement a prototype system of progressive profiling and investigate different methodology choices in the process. Our results show that by performing progressive profiling over multiple generations of a program, we can save a sizable amount of profiling time while sacrificing little profile quality.

Citations