Date of Award
Master of Applied Science (MASc)
Douglas G. Down
For the past few years, research in the area of computer clusters has been a hot topic. The main focus has been towards on how to achieve the best performance in such systems. While this problem has been well studied, many of the solutions maximize performance at the expense of increasing the amount of power consumed by the cluster and consequently raising the cost of power usage. Therefore, power management (PM) in such systems has become necessary. Many PM policies are proposed in the literature to achieve this goal for both homogeneous and heterogeneous clusters.
In this work, in the case of homogeneous clusters, we review two applicable policies that have been proposed in the literature for reducing power consumption. We also propose a power saving policy, that uses queueing theory formulas, which attempts to minimize power consumption while satisfying given performance constraints. We evaluate this policy by using simulation and compare it to other applicable policies.
Our main contribution is for heterogeneous clusters. We suggest a task distribution policy in order to reduce power consumption. Our suggested policy requires solving two linear programming problems (LPs). Our simulation experiments show that our proposed policy is successful in terms of achieving a significant power savings in comparison to other distribution policies, especially in the case of highly heterogeneous clusters.
AL-DAOUD, HADIL, "POWER-AWARE SCHEDULING FOR SERVER CLUSTERS" (2010). Open Access Dissertations and Theses. Paper 4101.
McMaster University Library