Programming for Malleability with Hybrid MPI-2 and OpenMP – Experiences with a Simulation Program for Global Water Prognosis

This paper reports on our experiences in parallelizing WaterGAP, an originally sequential C++ program for global assessment and prognosis of water availability. The parallel program runs on a heterogeneous SMP cluster and combines different parallel programming paradigms: First, at its outer level, it uses master/slave communication implemented with MPI. Second, within the slave processes, multiple threads are spawned by OpenMP directives to exploit data parallelism. Time measurements show that the hybrid scheme pays off. It adapts to the heterogeneity of the cluster by using multiple threads only for the largest tasks and mapping these to multiprocessor nodes. Third, the program is malleable, which has been accomplished with the dynamic process management facilities of MPI-2, based on the MPICH2 implementation. In particular, it is possible to increase the number of processes while the program is running. Malleability is an important feature in both batch systems and grid environments. We discuss the support that MPI-2 provides for malleability.