Power Control for GPU Clusters in Processing Large-scale Streams

Many emerging online data analysis applications require Large-scale streams data processing. GPU cluster is becoming a significantly parallel computing scheme to handling large-scale streams data tasks. However power optimization is a challenging issue. In this paper, we present a novel power consumption control model to shift power budge among nodes in the cluster based on their real workload needs, while capping redundancy energy and controlling the total power budge of the cluster to keep or below a constraint imposed by its power supplies. Our controller is very suitable to the dynamic workloads task model and designed based on an Multi-Input_Multi-Output control theory. We analyze the power consumption behaviors of GPU cluster and the variation of workload. The detailed control problem formulation is presented and analyzed in theory. We finally conduct simulation experiments on a physical cluster to compare our controller with two state-of-the-art controllers. The experimental results demonstrate that our proposed controller outperforms the other controllers by having more accurate control and more stability.