Online Demand Response of GPU Cloud Computing with DVFS

Abstract

GPU cloud computing is emerging as a new type of cloud service that drives computation-extensive jobs, such as big data analytics and distributed machine learning. The introduction of GPU brings parallel processing power at the cost of excessive energy consumption. Dynamic Voltage and Frequency Scaling (DVFS) is a promising method to control energy consumption of GPU VMs. This work focuses on using DVFS to reduce energy of cloud computing in datacenter demand response. We first consider an online demand response scenario where users arrive stochastically, aiming at maximizing social welfare and meeting energy reduction goals by employing DVFS. We address the challenge posed by DVFS through a new technique of compact infinite optimization. A more practical scenario where both energy and resource limitations present is further studied. We design a primal-dual approximation algorithm that can compute a feasible solution in polynomial time with guaranteed approximation ratio, and a payment scheme that works in concert to form a truthful cloud job auction.