Thermal-aware job allocation and scheduling for three dimensional chip multiprocessor

Abstract

In this paper, we propose a thermal-aware job allocation and scheduling algorithm for three-dimensional (3D) chip multiprocessor (CMP). The proposed algorithm assigns hot jobs to the cores close to the heat sink and cool jobs to the cores far from the heat sink, subject to thermal constraints. The direct effect of the proposed algorithm on a 3D-CMP system is that, the heat from hot jobs is removed off the chip faster than the temperature-aware methods. Therefore we are able to keep the chip cooler and in better thermal condition. Experimental results show that, comparing to the temperature-aware method, our algorithm achieves: 1) less hot spots; 2) better performance; 3) smaller temporal temperature variation; 4) lower peak temperature. The proposed algorithm reduces hot spots by more than 95% when workload contains cool jobs; and by 36% when workload does not contain cool jobs. It also boosts the system performance by 5% on average under various workloads. The temporal temperature variation is reduced by 60% and its standard deviation is decreased by 50%. In addition, the proposed algorithm achieves 1.8°C ∼5°C reduction in peak temperature.