An MPI driven parallelization strategy for different computing platforms: A case study

In this work we show how the MPI collective communication primitives can offer precious guidelines in the parallelization of a given sequential application. We explore a parallelization strategy that starting by a real sequential algorithm (the well known N_body algorithm), drives the programmer to an effective parallel version by means of consecutive simple actions (even if automated), that allows a great reuse of the sequential code without substantial changes in the original algorithm structure. The performance issues of the parallelization activity have been analyzed using two distributed architectures, the first one consisting of all homogeneous nodes and the latter characterized by heterogeneous computational resources. The load sharing of the application needed to match the heterogeneity of the target machine can be obtained once again in a straightforward manner.