Collaborative compilation

Abstract

Modern optimizing compilers use many heuristic solutions that must be tuned empirically. This tuning is usually done "at the factory" using standard benchmarks. However, applications that are not in the benchmark suite will not achieve the best possible performance, because they are not considered when tuning the compiler. Collaborative compilation alleviates this problem by using local profiling information for at-the-factory style training, allowing users to tune their compilers based on the applications that they use most. It takes advantage of the repeated compilations performed in Java virtual machines to gather performance information from the programs that the user runs. For a single user, this approach may cause undue overhead; for this reason, collaborative compilation allows the sharing of profile information and publishing of the results of tuning. Thus, users see no performance degradation from profiling, only performance improvement due to tuning. This document describes the challenges of implementing collaborative compilation and the solutions we have developed. We present experiments showing that collaborative compilation can be used to gain performance improvements on several compilation problems. In addition, we relate collaborative compilation to previous research and describe directions for future work.