Effectiveness of Neural Networks for Power Modeling for Cloud and HPC

Abstract

Power consumption of servers and applications are of utmost importance as computers are becoming ubiquitous, from smart phones to IoT and full-fledged computers. To optimize their power consumption, knowledge is necessary during execution at different levels: for the Operating System to take decisions of scheduling, for users to choose between different applications. Several models exist to evaluate the power consumption of computers without relying on actual wattmeters: Indeed, these hardware are costly but also usually have limits on their pooling frequency (usually a one-second frequency is observed) except for dedicated professional hardware. The models link applications behavior with their power consumption, but up to now there is a 5% wall: Most models cannot reduce their error under this threshold and are usually linked to a particular hardware configuration. This article demonstrates how to break the 5% wall of power models. It shows that by using neural networks it is possible to create models with 1% to 2% error. It also quantifies the reachable precision obtainable with other classical methods such as analytical models.