Decentralised grid scheduling approach based on multi-agent reinforcement learning and gossip mechanism

Abstract

As an important class of resource allocation approaches, decentralised job scheduling in large-scale grids has to deal with the difficulties in acquiring timely model information and improving performance by autonomous coordination. In this study, a gossip-based reinforcement learning (GRL) method is proposed for decentralised job scheduling in grids. In the GRL method, a decentralised scheduling architecture based on multi-agent reinforcement learning is presented to improve the scalability and adaptability of job scheduling. A gossip mechanism is designed to realise autonomous coordination among the decentralised schedulers. Simulation results show that the proposed GRL-based schedulers can complete the task of grid job scheduling effectively and achieve load balancing efficiently.