MPICH/MADIII : a cluster of clusters enabled MPI implementation

Abstract

This paper presents an MPI implementation that allows an easy and efficient use of the interconnection of several clusters, of potentially heterogeneous nature (as far as the network is concerned). We describe the underlying communication subsystem used. The mechanisms within MPI that inform the user of the underlying topological structure are detailed The performance figures obtained with this MPI implementation are discussed and advocate for the use of such a solution on this particular type of architecture.