Techniques for global optimization of message passing communication on unreliable networks

Abstract

In this paper, we present techniques to improve the performance of parallel and distributed applications running on distributed systems built with unreliable local-area networks. The optimizations tailor the message passing system used by the application to the communication pattern exhibited by the application. The optimizations are global and application dependent since communication patterns vary from application to application. The techniques improve both the execution times and scalability of many parallel applications as well as distributed system services.