Overlapping communication and computation by using a hybrid MPI/SMPSs approach

Abstract

Communication overhead is one of the dominant factors that affect performance in high-performance computing systems. To reduce the negative impact of communication, programmers overlap communication and computation by using asynchronous communication primitives. This increases code complexity, requiring more effort to write parallel code and making less readable code. This paper presents the hybrid use of MPI and SMPSs (SMP superscalar), a task-based shared-memory programming model, enhanced with a restart mechanism allowing the programmer to introduce the asynchronism that is necessary to enable the effective communication/computation overlap in a productive way. We demonstrate the hybrid use of MPI/SMPSs with the high-performance LINPACK benchmark, which uses the lookahead technique to overlap communication and computation. MPI/SMPSs improves the performance of a pure MPI with look-ahead by 7,6% on a 1024 processors machine. In addition to better performance, hybrid MPI/SMPSs substantially reduces code complexity, it is less sensitive to network bandwidth and operating system noise, and improves the use of main memory.