Load Balancing for Parallel Discrete Event Simulation of Stochastic Reaction and Diffusion

Load balancing plays an essential role in parallel and distributed computing. In this paper, we present the load balancing issue of our multi-threaded parallel discrete event simulator for stochastic simulation and our solution. The computational and communicational workload of each thread is assessed during simulation, and the state of the simulation is represented by this assessment. We utilize a multi-state Q-learning approach to determine the parameters of workload migration-the type of objective workload, the number of threads involved in migration and the load to be exchanged. Workload is transferred within the same process by mapping it to the new thread directly and by communication between processes. As the test on a calcium wave model shows, our solution achieved 31% improvement in execution time by migrating workload among threads within the same process, 21% improvement when workload is transferred within the same node, and 16% when remote communication is used.