DeepScheduling: Grid Computing Job Scheduler Based on Deep Reinforcement Learning

Abstract

Grid systems are large-scale platforms which consume a considerable amount of energy. Several efficient resource/power management strategies were proposed by the specialized literature. However, most of the proposed strategies are rule-based policies which do not exploit workload patterns. Deploying the same set of rules on systems using different usage patterns, and platform settings, may lead to a sub-optimized setup. Due to the complex nature of grid systems, tailoring such a system-specific policy is not a straightforward task. In this paper, we explore a Deep Reinforcement Learning (DRL) method to build an adaptive energy-aware scheduling policy. We trained our algorithm using real workload traces from Grid’5000 platform. Our experiments pointed out an energy setup saving up to 7%, as well as average requests waiting time reduction of 27%. Finally, the resuslts clarify the importance of explore the workload to build system-specific policies.