Comparison of Message Aggregation Strategies for Parallel Simulations on a High Performance Cluster
Abstract
Parallel simulations of fine grain applications usually generate a large amount of messages. The overhead for sending these messages over an interconnection network can dramatically limit the speedup of a parallel simulation. In this case, message aggregation techniques can increase the granularity of the application and reduce the communication overhead. This paper compares sender-initiated and receiver-initiated message aggregation strategies and addresses their performance when applied to parallel simulations of high-speed communication network models on a high performance cluster. Receiver-initiated strategies were initially motivated by the cyclic behavior of the parallel simulation kernel and by the capabilities of the test-bed environment to handle large messages up to 2000 bytes with small overheads.