Automatic performance debugging of SPMD-style parallel programs

摘要: Automatic performance debugging of parallel applications includes two main steps: locating bottlenecks and uncovering their root causes for optimization. Previous work fails to resolve this challenging issue in ways: first, several previous efforts automate bottlenecks, but present results a confined way that only identifies problems with priori knowledge; second, tools take exploratory or confirmatory data analysis automatically discover relevant relationships, these do not focus on causes. The simple program multiple (SPMD) programming model is widely used both high computing Cloud computing. In paper, we design implement an innovative system, AutoAnalyzer, automates the process SPMD-style programs, including collection, behavior analysis, AutoAnalyzer unique terms features: without any prior knowledge, it locates uncovers optimization; lightweight size be collected analyzed. Our contributions are three-fold: propose effective clustering algorithms investigate existence cause dissimilarity code region disparity, respectively; meanwhile, searching locate bottlenecks; basis rough set theory, approach uncover third, cluster systems different configurations, use production applications, written Fortran 77, one open source code-MPIBZIP2 (http://compression.ca/mpibzip2/), C++, verify effectiveness correctness our methods. For three also experimental investigating effects metrics bottlenecks.