PASS: Power-Aware Scheduling of Mixed Applications with Deadline Constraints on Clusters
Cong Liu, Xiao Qin†, and Shuang Li
Department of Computer Science and Software Engineering
Auburn University, Auburn, AL 36849
Reducing energy consumption has become a pressing issue in cluster computing systems not only for minimizing electricity cost, but also for improving system reliability. Therefore, it is highly desirable to design energy-efficient scheduling algorithms for applications running on clusters. In this paper, we address the problem of non-preemptively scheduling mixed tasks on power-aware clusters. We developed an algorithm called Power Aware Slack Scheduler (PASS) for tasks with different priorities and deadlines. PASS attempts to minimize energy consumption in addition to maximizing the number of tasks completed before their deadlines. To achieve this goal, high-priority tasks are scheduled first in order to meet their deadlines. Moreover, PASS explores slacks into which low-priority tasks can be inserted so that their deadlines can be guaranteed. The dynamic voltage scaling (DVS) technique is used to reduce energy consumption by exploiting available slacks and adjusting appropriate voltage levels accordingly. Simulation results demonstrate that compared with a well-known energy-efficient algorithm - CC-EDF, PASS saves up to 60 percent of energy dissipation. With respect to the number of high-priority tasks meeting deadlines, PASS outperforms the existing approach by over 10 percent without degrading the overall performance. PASS successfully schedules tasks with hard deadlines in a mix of tasks with soft deadlines. In doing so, PASS embraces a new feature that allows clusters to support a variety of real-time applications, making clusters amenable for commercialization.
This paper appeared in the Proceedings of the 17th IEEE International Conference on Computer Communications and Networks (ICCCN), St. Thomas, Virgin Islands, Aug. 2008.
†Corresponding author. http://www.eng.auburn.edu/~xqin
The work reported in this paper was supported by the US National Science Foundation under Grants No. CCF-0742187, No. CNS-0757778, No. CNS-0831502, No. OCI-0753305, No. DUE-0621307, and No. DUE-0830831, and Auburn University under a startup grant.