Observations on MPI-2 support for hybrid master/slave applications in dynamic and heterogeneous environments

Abstract

Large-scale MPI programs must work with dynamic and heterogeneous resources. While many of the involved issues can be handled by the MPI implementation, some must be dealt with by the application program. This paper considers a master/slave application, in which MPI processes internally use a different number of threads created by OpenMP. We modify the standard master/slave pattern to allow for dynamic addition and withdrawal of slaves. Moreover, the application dynamically adapts to use processors for either processes or threads. The paper evaluates the support that MPI-2 provides for implementing the scheme, partly referring to experiments with the MPICH2 implementation. We found that most requirements can be met if optional parts of the standard are used, but slave crashes require additional functionality. © Springer-Verlag Berlin Heidelberg 2006.