LAUSR

This paper proposes efficient strategies that shave Data Centers (DCs)’ monthly peak power demand with the aim of reducing the DCs’ monthly expenses. Specifically, the proposed strategies allow to decide: $i)$i) when and how much of the DC's workload should be delayed given that the workload is made up of multiple classes where each class has a certain delay tolerance and delay cost, and $ii)$ii) when and how much energy should be charged/discharged into DCs’ batteries. We first consider the case where the DC's power demands throughout the whole billing cycle are known and present an optimal peak shaving control strategy for it. We then relax this assumption and propose an efficient control strategy for the case when (accurate/noisy) predictions of the DC's power demands are only known for short durations in the future. Several comparative studies based on real traces from a Google DC are conducted in order to validate the proposed techniques.