A multi-agent architecture for scheduling of high performance services in a GPU cluster