Prediction of communication latency over complex network behaviors on SMP clusters

Abstract

Using MPI as communication interface, one or several applications may introduce complex communication behaviors over the network of a cluster. This effect is increased when nodes of the cluster are multi-processors, and where communications can income or outgo from the same node with a common interval time. Our interest is to understand the effects of complex communication schemes over a network of SMP nodes, and then to identify different network conflicts. Network conflicts stand for a network component (links, NIC, or MPI stack, etc.) shared between communications. This paper describes a set of experiments generating network conflicts and their analysis, in order to accurately predict communication latencies. This analysis is based on cluster of bi-processor with a Myrinet network, using as communication interface a LAM MPI-2 implementation over GM protocol. © Springer-Verlag Berlin Heidelberg 2005.