LAUSR

To reduce the ecological impact of datacenters and to operate them in areas with unreliable electrical grid, the use of on-site renewable energy sources (RESs) is actively studied nowadays. The main barrier to their wide adoption is related to the intermittent nature of common RESs, such as solar panels or wind turbines. Several works already demonstrated the suitability of energy storage devices and energy consumption scheduling to mitigate this issue. We propose to abstract such an infrastructure by two independent black-box systems: A power producer takes care of the RESs and storage devices; a power consumer manages the datacenter itself. An optimization module cooperates with these two systems to find the best power production and consumption plans for an upcoming time window. Each system having different goals, we tackle the multi-objective problem using a new evolutionary algorithm to find a set of ideal trade-offs. Evaluating the outcome of a power plan for an electrical or computing system is however usually costly. To reduce the number of evaluations of these objective functions, a common method is to use a cheap surrogate. Using time series properties, a novel surrogate method is presented. The performance of our approach is evaluated with maximization of quality of service and minimization of greenhouse emissions as respective objectives. Experiments are performed first using a simplified datacenter model with computable Pareto front, with the main findings replicated in a realistic model and a datacenter scheduler from the literature.