A Hybrid MPI-OpenMP Implementation of the Conjugate Gradient Method in Java.

Abstract

The message passing paradigm is the dominating programming model for parallel processing on distributed-memory architectures. OpenMP is the most widely used parallel programming model for shared-memory systems. Contemporary parallel computers consist of multiple shared-memory computers connected via a network so that a hybrid parallelization approach combining message passing and OpenMP is often desired. We present a hybrid parallelization of an iterative method for the solution of linear systems with a large sparse symmetric coefficient matrix which is implemented in Java. To this end, we bring together mpiJava and JOMP and compare the performance on a Sun Fire 6800 system with a hybrid parallel version of the same problem implemented in Fortran. The scalability of both implementations is similar while, surprisingly, the performance of the Fortran version is only a factor of about 2 larger than the Java version when up to 22 processors are used.