Assessing contention effects of all‐to‐all communications on clusters and grids

Abstract

One of the most important collective communication patterns used in scientific applications is the complete exchange, also called All-to-All. Although efficient algorithms have been studied for specific networks, general solutions like those available in wellknown MPI distributions (e.g. the MPI_Alltoall operation) are strongly influenced by the congestion of network resources. In this paper we present an integrated approach to model the performance of the Allto-All collective operation, which consists in identifying a contention signature that characterizes a given network environment, using it to augment a contentionfree communication model. This approach, assessed by experimental results, allows an accurate prediction of the performance of the All-to-All operation over different network architectures with a small overhead. We also discuss the problem of network contention in a grid environment, studying some strategies to minimize the impact of contention on the performance of an Allto-All operation.