Stochastic Thermal Control of a Multicore Real-Time System

Abstract

This paper deals with thermal management of a multicore processor executing multiple stochastic real-time job streams. The main objective is to reduce the chip-wide temperature gradient to decelerate processor aging, and the subordinate goal is to decrease the hotspot temperature. A pair of active and passive cores is dedicated to each stream, which the active one services the corresponding real-time jobs. In order to reduce the chip-wide temperature gradient between cores, the active and passive cores of an individual stream are replaced at appropriate times through job migration. The thermal management of this system is a specific stochastic control problem. Regarding the inter-effects of core temperatures and the stochastic nature of the system, systematic achievement of the objective needs an appropriate method. The control theory of Markov jump linear system (MJLS) has been used to design the desired thermal controller and analytically study its stability. The efficacy of the proposed approach in terms of the thermal management objectives is investigated through simulation experiments.