Impact of Message-Ordering in Wormhole-Routed Multicomputers

This paper analyzes the impact of message-ordering, between outgoing messages from a sender to multiple receivers (called multicasts), on the completion time of a program for distributed-memory systems. We study how best to order a set of outgoing messages by taking into account message criticality and architectural issues like multiple injection ports, routing adaptivity, and link contention. First, the simple message-ordering algorithm of Dikaiakos et al. 8], proposed only for fully-connected one-port system, is enhanced to obtain a static message-ordering algorithm. This enhanced algorithm is based on message criticality obtained from the temporal dependencies of the program and works for non-fully-connected systems. Next, a dynamic message-ordering algorithm is proposed which additionally considers architectural issues including link contention and multiple ports for wormhole-routed systems. Both algorithms are compared against naive sequential message-ordering on an event-driven simulator for Gaussian Elimination and random task graphs. Simulation experiments show a reduction in program completion time up to 33.1% by using static over sequential ordering. Similarly, the dynamic ordering shows additional improvement of up to 17.7% over static ordering. This research indicates that good message-ordering schemes need to be used as the last step in the mapping and scheduling process to generate eecient programs for wormhole-routed distributed-memory systems.