FOOD-WEB SIMULATION ON PARALLEL COMPUTERS - INTERPROCESSOR COMMUNICATION BENCHMARKS

A major problem in simulating ecological systems on distributed memory multiprocessors is the cost of communicating states of variables among processors that require those states to update other variables. I measured the communication cost between two INMOS Transputers organized in a task-farm topology when simulating artificial food-webs with Lotka-Volterra dynamics. When processor communication effort is minimized by transferring only the variables required by each processor, speed is increased by only 5-9% compared with transferring all variables. The overall speed-up obtained by using two processors compared to one depends on the number of system nodes in the model, but was 1.9 when 96 nodes were simulated. When the step-size of the Euler integration method is adjusted to produce dynamics quantitatively similar to Runge-Kutta, the Euler method is slower despite reduced inter-processor communication overhead produced by fewer calculations of derivatives required for reach variable update.