Automatic translation of non-repetitive OpenMP to MPI

Fahed A Jubair, Purdue University

Abstract

Cluster platforms with distributed-memory architectures are becoming increasingly available low-cost solutions for high performance computing. Delivering a productive programming environment that hides the complexity of clusters and allows writing efficient programs is urgently needed. Despite multiple efforts to provide shared memory abstraction, message-passing (MPI) is still the state-of-the-art programming model for distributed-memory architectures. Writing efficient MPI programs is challenging. In contrast, OpenMP is a shared-memory programming model that is known for its programming productivity. Researchers introduced automatic source-to-source translation schemes from OpenMP to MPI so that programmers can use OpenMP while targeting clusters. Those schemes limited their focus on OpenMP programs with repetitive communication patterns (where the analysis of communication can be simplified). This dissertation reduces this limitation and presents a novel OpenMP-to-MPI translation scheme that covers OpenMP programs with both repetitive and non-repetitive communication patterns. We target laboratory-size clusters of ten to hundred nodes (commonly found in research laboratories and small enterprises). With our translation scheme, six non-repetitive and four repetitive OpenMP benchmarks have been efficiently scaled to a cluster of 64 cores. By contrast, the state-of-the-art translator scaled only the four repetitive benchmarks. In addition, our translation scheme was shown to outperform or perform as well as the state-of-the-art translator. We also compare the translation scheme with available hand-coded MPI and Unified Parallel C (UPC) programs.

Degree

Ph.D.

Advisors

Eigenmann, Purdue University.

Subject Area

Computer Engineering

COinS