University of Houston

OpenMP is an application programming interface (API) for parallel programming on shared memory systems. The introduction of OpenMP tasks allows programmers to express unstructured parallelism at a high level of abstraction and make the OpenMP runtime responsible about the burden of scheduling parallel execution. The ability to observe performance and scheduling issues among OpenMP programs has been a challenge due to the lack of performance interface standards in the runtime layer. In this work, we present a methodology for developing a similarity-based system using our proposed Collector API extensions to the OpenMP Runtime API (ORA), used for OpenMP profiling. We exploit this system to adaptively predict the OpenMP scheduling strategy that should distinctly be applied to each subset of OpenMP programs to obtain the best performance with a focus on applications exhibiting irregular parallelism. We evaluate the usefulness of the subsets of programs obtained using a proposed statistical runtime analysis model by demonstrating that the average cache miss rate, speedup, translation lookaside buffer misses, etc of the entire subset of programs can be predicted with a reasonable accuracy. This requires (1) a powerful performance framework to analyze OpenMP programs and scheduling strategies with respect to exploiting data locality, maintaining load balance, and minimizing overhead costs, (2) a reliable statistical model that can be used to measure similarity between OpenMP programs exhibiting unstructured behavior, and (3) a detailed evaluation and validation study against a broad range of OpenMP applications.