Edge and Cloud Provider Cost Minimization by Exploiting Extended Voltage and Frequency Margins

Abstract

Energy costs contribute significantly to the total cost of operation for cloud and edge infrastructure providers. Both conventional (Voltage and Frequency Scaling) and more aggressive (undervolting, overclocking) techniques can be applied to reduce the energy footprint of the infrastructure and thus the cost of the operator. However, these techniques may affect the quality of service (QoS), potentially activating service level agreement (SLA) violation penalties for the provider. In this paper we model and study this tradeoff. We find that undervolting is the most effective of the three techniques in reducing infrastructure operation cost. We identify optimal operating points, and we study the effect of different parameters, such as the severity of SLA penalties, the length of the job and the existence of error protection mechanisms, on the optimal operating point and the extent of potential benefits. © 2018 The authors and IOS Press.