Optimizing function placement for large-scale data-center applications

摘要

Modern data-center applications often comprise a large amount of code, with substantial working sets, making them good candidates for code-layout optimizations. Although recent work has evaluated the impact of profile-guided intramodule optimizations and some cross-module optimizations, no recent study has evaluated the benefit of function placement for such large-scale applications. In this paper, we study the impact of function placement in the context of a simple tool we created that uses sample-based profiling data. By using sample-based profiling, this methodology follows the same principle behind AutoFDO, i.e. using profiling data collected from unmodified binaries running in production, which makes it applicable to large-scale binaries. Using this tool, we first evaluate the impact of the traditional Pettis-Hansen (PH) function-placement algorithm on a set of widely deployed data-center applications. Our experiments show that using the PH algorithm improves the performance of the studied applications by an average of 2.6%. In addition to that, this paper also evaluates the impact of two improvements on top of the PH technique. The first improvement is a new algorithm, called C3, which addresses a fundamental weakness we identified in the PH algorithm. We not only qualitatively illustrate how C3 overcomes this weakness in PH, but also present experimental results confirming that C3 performs better than PH in practice, boosting the performance of our workloads by an average of 2.9% on top of PH. The second improvement we evaluate is the selective use of huge pages. Our evaluation shows that, although aggressively mapping the entire code section of a large binary onto huge pages can be detrimental to performance, judiciously using huge pages can further improve performance of our applications by 2.0% on average.