Mise en place et évaluation d'un algorithme de répartition de charge pour les plateformes de simulations distribuées basées sur les systèmes multiagents

This paper discusses load balancing in Multi-Agent Systems (MAS) through a generic algorithm based on agent’s distribution. The need is born from the observation of frequent load-balancing problems over simulations based on multiagent systems. Miro which is a software system developed for scalable urban mobility simulation is a concrete example of this type of problems. The difficulty of this work lies in the consideration of simulations specificities: autonomy of distributing entities, high unpredictability of the system...We adapt a load balancing algorithm called “Comet” to specificities of distributed simulations based on multi-agent systems. This algorithm proposes a credit-based load balancing policy which evaluates agent affinity to a machine and determines which the best candidates for migration are. Besides, this paper presents an implementation and an evaluation of the algorithm on a simulator developed with NetLogo. The final purpose is to identify which parameters are suitable to the best running of this load-balancing algorithm on simulations based on multi-agent systems.