Power-aware Work Stealing in Homogeneous Multicore Systems

摘要

Excessive power consumption affects the reliability of cores, requires expensive cooling mechanisms, reduces battery lifetime, and causes extensive damage to the device. Hence, managing the power consumption and performance of cores is an important aspect of chip design. This research aims to achieve efficient multicore power monitoring and control via operating system based power-aware task scheduling. The main objectives of power-aware scheduling are: 1) lowering core's power consumption level, 2) maintaining the system within an allowable power envelope, and 3) balancing the power consumption across cores; without significant impact on time performance. In previous research we have explored power-aware task scheduling at the single core level referred to as intra-core scheduling. This paper reports on a research on a power-aware form of inter-core scheduling policy referred to as work stealing. Work stealing is a special case of task migration, where a "starving" core attempts to steal tasks from a "victim", i.e., a "loaded" core. We have performed experiments with ten variants of the work stealing that consider both the power and the performance attributes of the system in the process of selecting a victim core. The experiments conducted show that the power-aware inter-core stealing policies have high potential for power efficient task scheduling with tolerable effect on performance.