Load Balancing for Distributed Battlefield Simulations: Tradeoffs in Workload and Communications

Load balancing attempts to optimize the utilization of processors in a parallel computing systems, and dynamic load balancing is a prime candidate for improving the performance of distributed battlefield simulation systems. Last year we reported on development of test-bed developed to assist in the empirical exploration of a number of dynamic load balancing heuristics. Unique to the test-bed was a modification of the classical discrete-event simulation (DES) scheduling paradigm that enabled us to determine processor load at the application level. Experimental runs considered a number of load-balancing heuristics, corroborated results reported by other researchers, and provided confidence that our approach is indeed feasible. Absent from this initial study was a consideration of cost measures for the system and a recognition of the conflict between distributing workload evenly and minimizing communication costs. For a load balancing system to be effective, the cost of balancing load must be less than the cost of the status quo. The cost is manifested by the monitoring, selection, transport, and initialization time, versus the processing and bandwidth requirements. Without proper monitoring and calibration, it possible to spend more time trying to balance the load than it does actually processing productive work. In this paper we present the implementation of additional test-bed infrastructure designed to capture these tradeoffs. Moreover, we motivate the selection of heuristics to be considered, present the results of experimental runs with these heuristics, and discuss the implications of the results.