Towards an energy-efficient Data Center Network based on deep reinforcement learning

Data Center Network (DCN) plays a crucial role in orchestrating the physical or virtual resources in data centers to meet the requirements of Internet of Things and Cloud Computing. The energy efficiency should be seriously considered for DCNs with large-scale switch devices which support numerous realtime network flow demands, especially for enormous flow demands from IoT devices. Typically, the network energy conservation can be achieved by optimizing routing and flow scheduling with energy awareness, targeting at powering off as many idle and low-loaded network devices as possible. For energy efficiency objective, in this paper we address a combinatorial optimization problem, named Multi-Commodity Flow (MCF) problem which optimizes the bandwidth allocation and routing to reduce the energy consumption. We propose a framework which has the lookahead ability of predicting flow demands in DCNs to dynamically feed the MCF problem as inputs. A Long Short-Term Memory (LSTM) network is exploited for flow demand prediction in DCNs and a Deep Reinforcement Learning (DRL) algorithm is tailored for solving the MCF problem. In experiments, we evaluate the predicted flow demands which simulate real flow demands and conduct a comparison between our DRL scheme with the baseline and optimizer to show the advantage of the DRL solution in optimality and efficiency. © 2022 Elsevier B.V.