Near–Critical Path Analysis: A Tool for Parallel Program Optimization

Program activity graphs (PAGs) can be constructed from timestamped traces of appropriate execution events. Information about the activities on the k longest execution paths is useful in the analysis of parallel program performance. In this paper, four algorithms for finding the near–critical paths of PAGs are compared, including a best–first search (BFS) algorithm that is worst–case asymptotically optimal in terms of both time and space. Results confirming the practical efficiency of the BFS algorithm are presented for five application programs. A framework for using the near–critical path information is also described. The framework includes statistical summaries and visualization capabilities that build upon the foundation of existing performance analysis tools. Within the framework, guidance is provided by the Maximum Benefit Metric, which uses near–critical path data to predict the maximum overall performance improvement that may be realized by optimizing particular critical path activities. The framework is validated by a case study: development and tuning of a parallel near–critical path algorithm. Examples are cited illustrating the utility of both framework components. Index Terms — Parallel computing, performance analysis, performance metric, instrumentation, program activity graph, critical path, longest path algorithm.