Predicting Contention in Distributed-Memory Machines

Arjan J.C. van Gemund [email protected] Department of Electrical Engineering Delft University of Technology P.O.Box 5031, NL-2600 GA Delft, The Netherlands Abstract A compile-time prediction technique is outlined that yields low-cost, highly symbolic performance models, to be used during the initial optimization loops in parallel system design. Aimed to provide an acceptable accuracy across a large parameter search space the approach is based on extending conventional static analysis with asymptotic queueing analysis in order to account for potentially dominating e ects of resource contention. In this paper we report on the accuracy of the prediction method when compared to simulation results as well as compared to actual measurement results on a distributed-memory machine.