Built-in Stigmergy-Based Load Balancing Model for HPC Clusters

Abstract

HPC load balancing aims to manage and share multiple computational resources in order to increase system performance. Several researches have been conducted on the issue of HPC load balancing. Their common point is that they consider processes (workload) as passive units to be balanced among cluster nodes (computational resources) to ensure optimal load balancing. The hypothesis behind our work assumes that the unit to be balanced, i.e. a process (computational task), is more likely to know its needs, and thus more likely to decide whether to migrate to another node or not. Therefore, the load balancing issue is studied at the parallel programming phase. Generated processes, after being recompiled using our LB integrator tool, will be systematically granted a built-in intelligence. The latter enables them to monitor their environment and migrate from one node to another in case of overloading. For this end, we have designed a built-in load balancing model based on stigmergy and collective intelligence concepts (BSLB). A novel P2P cluster computing system has been designed using Repast toolkit and the BSLB algorithm, allowing arbitrary users to initiate intelligent processes.