Autoscaling Effects in Speed Scaling Systems

Abstract

In this paper, we study the dynamics of coupled speed scaling systems, in which service rate is a function of system occupancy. We focus on both Processor Sharing (PS) and Shortest Remaining Processing Time (SRPT) as scheduling disciplines, and study their speed scaling dynamics under heavy load. Using a combination of Markov chain analysis and discrete-event simulation, we identify several important properties of speed scaling systems, which we call the autoscaling effect, the α effect, and the saturation effect. We also identify different overload regimes for PS and SRPT. In particular, SRPT exhibits a starvation effect that differs from the compensation effect of PS. These dynamics lead to different stability, fairness, and robustness properties for PS and SRPT under heavy load.