Ring Algorithms on Heterogeneous Windows-based Clusters with Various Message Passing Environments

The aim of the present work is the development of efficient and self-adaptive ring algorithms on heterogenous Windows-based clusters. We show that a virtual ring of processes, with a number of processes on each node proportional to its relative speed, greatly reduces load imbalance and allows to achieve good performance even on highly heterogeneous systems. As test application we consider the computation of long- and short-range interactions. Two different implementations of MPI for Windows are considered (MPICH and MPICH2) and some comparisons with PVM are also performed. The analysis is quite general and can be applied to similar problems. From the algorithm analysis we obtain both a full computer simulator of ring applications and some simplified indices of performance, useful to quickly adapt the application to a given platform.