Transforming blocking MPI collectives to Non-blocking and persistent operations

This paper describes Petal, a prototype tool that uses compiler-analysis techniques to automate code transformations to hide communication costs behind computation by replacing blocking MPI functions with corresponding nonblocking and persistent collective operations while maintaining legacy applications' correctness. In earlier work, we have already demonstrated Petal's ability to transform point-to-point MPI operations in complement to the results shown here. The contributions of this paper include the approach to collective operation transformations, a description of achieved functionality, examples of transformations, and demonstration of performance improvements obtained thus far on representative sample MPI programs. Depending on system scale and problem size, the transformations yield a speedup of up to a factor of two. This tool can be used to transform useful classes of new and legacy MPI programs to use the newest variants of MPI functions to improve performance without manual intervention for forthcoming HPC systems and updated versions of the MPI standard.