Dynamic load balancing in multi-agent spatial simulation

Abstract

This paper presents dynamic load balancing in a parallelizing library for multi-agent spatial simulation (named MASS). Our load-balancing algorithms calculate per-thread CPU load right after every function call and adjust a data size to be passed to each thread for the next function call. We implemented three different thread-based load-balancing algorithms, each using (1) an entire history, (2) a recent time window and (3) a slope of the CPU loads. The paper presents our implementation of these three algorithms in MASS as well as performance evaluation with two multithreaded applications: Wave2D and SugarScape. Furthermore, to demonstrate the slope-based algorithm's superiority to the other two, we compared them over a cluster of computing nodes.