Multi-Agent Deep Reinforcement Learning Framework for Renewable Energy-Aware Workflow Scheduling on Distributed Cloud Data Centers

The ever-increasing demand for the cloud computing paradigm has resulted in the widespread deployment of multiple datacenters, the operations of which consume very high levels of energy. The carbon footprint resulting from these operations threatens environmental sustainability while the increased energy costs have a direct impact on the profitability of cloud providers. Using renewable energy sources to satisfy the energy demands of datacenters has emerged as a viable approach to overcome the aforementioned issues. The problem of scheduling workflows across multi-cloud environments powered through a combination of brown and green energy sources includes multiple levels of complexities. First, the general case of workflow scheduling in a distributed system itself is NP-hard. The need to schedule workflows across geo-distributed cloud datacenters adds a further layer of complexity atop the general problem. The problem becomes further challenging when the datacenters are powered through renewable sources which are inherently intermittent in nature. Consequently, traditional workflow scheduling algorithms and single-agent reinforcement learning algorithms are incapable of efficiently meeting the decentralized and adaptive control required for addressing these challenges. To this end, we have leveraged the recent advancements in the paradigm of MARL (Multi-Agent Reinforcement Learning) for designing and developing a multi-agent RL framework for optimizing the green energy utilization of workflow executions across multi-cloud environments. The results of extensive simulations demonstrate that the proposed approach outperforms the comparison algorithms with respect to minimizing energy consumption of workflow executions by 47% while also keeping the makespan of workflows in par with comparison algorithms. Furthermore, with the proposed optimizations, the multi-agent technique learnt 5 times faster than a generic multi-agent algorithm.