Dynamic load balancing for switch-based networks

Abstract

Recently, switch-based networks of workstations (NOWs) have been introduced as an alternative for traditional parallel computers. Although many dynamic load balancing algorithms have been developed for point-to-point networks (static networks), little progress has been made on the load balancing in switch-based networks (dynamic networks). Thus, in this paper, we propose a dynamic load balancing algorithm, called the Switch Walking Algorithm (SWA), suitable for switch-based networks. In SWA, each processor's load information is gathered to form global load information, which is then used for load balancing. SWA is compared to a previous algorithm, called the Tree Walking Algorithm (TWA), which has been applied to switch-based networks. Through analysis, we show that SWA requires less communication time for distribution of global load information and migrates fewer tasks than TWA. Also, we show, through the implementation of a Mandelbrot set generation program, that SWA achieves about 20% better performance than TWA on a system with 32 processing elements.