A Multi-Agent Learning Approach to Resource Sharing across Computing Clusters

Resource management in clusters traditionally uses centralized approaches, which restricts the cluster scale. To expand this limit, we develop a multi-agent approach to sharing resources across clusters in a decentralized manner. We organize shared clustered into an overlay network and formulate resource sharing in such a network as a distributed sequential resource allocation problem (DSRAP). We then propose a multi-agent reinforcement learning algorithm for each cluster to learn both local allocation decision policy and task routing policy so that clusters cooperatively allocate tasks and maximize the global utility of the system. Heuristic strategies are developed to speed up the learning in such a complex problem. We compare our approach with a centralized allocation approach that can generates optimal solutions in some cases. Experimental results show that our approach is very effective and even outperforms the centralized allocation approach in some cases where it does not generate optimal solutions.