Event driven power consumption optimization control model of GPU clusters

Reducing power consumption for GPU cluster in large-scale stream computing process can bring various benefits such as reducing operating costs and environmental effect. We formulate the problem of power consumption as a constrained optimization problem, minimizing power state of cluster nodes to reduce power consumption while guaranteeing system performance and reliability. The proposed control model based on Model Prediction Control is designed to make a comprehensive metric of GPU cluster achieve expected performance, energy efficiency and reliability. It is different from the previous models, which just consider power consumption as the sole control objective. The event-triggering mechanism is introduced to reduce control overhead. It successfully separates sampling cluster status signals from control model. So the controller needs not to periodically interrupt computing process to solve optimal solutions. Finally, we evaluate and compare this control model with the previous control model by using artificial and real-world workloads. The experimental results show that our proposed control model is able to outperform existing techniques.