Statistical Analysis of Latency Through Semantic Profiling

Abstract

Most software profiling tools quantify average performance and rely on a program's control flow graph to organize and report results. However, in interactive server applications, performance predictability is often an equally important measure. Moreover, the end user is often concerned with the performance of a semantically defined interval of execution, such as a request or transaction, which may not directly map to any single function in the call graph, especially in high-performance applications that use asynchrony or event-based programming. It is difficult to distinguish functionality that lies on the critical path of a semantic interval from other activity (e.g., periodic logging or side operations) that may nevertheless appear prominent in a conventional profile. Existing profilers lack the ability to (i) aggregate results for a semantic interval and (ii) attribute its performance variance to individual functions. We propose a profiler called VProfiler that, given the source code of a software system and programmer annotations indicating the start and end of semantic intervals of interest, is able to identify the dominant sources of latency variance in a semantic context. Using a novel abstraction, called a variance tree, VProfiler analyzes the thread interleaving and deconstructs overall latency variance into variances and covariances of the execution time of individual functions. It then aggregates latency variance along a backwards path of dependence relationships among threads from the end of an interval to its start. We evaluate VProfiler's effectiveness on three popular open-source projects (MySQL, Postgres, and Apache Web Server). By identifying a few culprit functions in these complex code bases, VProfiler allows us to eliminate 27%--82% of the overall latency variance of these systems with a modest programming effort.