Total Exchange Performance Modelling Under Network Contention

Abstract

One of the most important collective communication patterns for scientific applications is the many to many, also called complete exchange. Although efficient All-to-All algorithms have been studied for specific networks, general solutions like those found in well known MPI distributions are strongly influenced by the congestion of network resources. In this paper we present our approach to model the performance of the All-to-All collective operation. Our approach consists in identifying a contention factor that characterises the network environment, and using it to augment a contention-free communication model. This approach allows an accurate prediction of the performance of the All-to-All operation over different network environments with a small cost. Indeed, we demonstrate the accuracy of our approach by presenting our experiments with three different network environments, Fast Ethernet, Giga Ethernet and Myrinet.

Modeling Network Contention Effects on All-to-All Operations

“…Nevertheless, our preliminary experiences (described in detail in [4]), suggest that the overload of the receiver is not enough to induce additional resource contention. Thus, the performance slowdown observed during the execution of the All-to-All operation is almost exclusively due to the saturation of the network, which causes packet loss.…”