Flipping the priority: effects of prioritising HTC jobs on energy consumption in a multi-use cluster

Abstract

High Throughput Computing (HTC) through the use of volunteer computing provides a compelling opportunity to perform large volumes of computation without the need to invest in computational resources. This relies on the good will of computer owners who volunteer their computer idle time. In scenarios when there is contention the HTC system will relinquish the computer -- normally achieved through termination or suspension. This often leads to longer turn-around times for the HTC jobs and an increase in energy consumption when the work is restarted elsewhere. However, within large organisations this distinction of who is the owner of a computer and who should take priority is not always clear. If a user enters a large cluster room (of identical computers) should they be allowed to use a computer which is servicing a HTC job when other computers are idle? Should HTC jobs be able to delay the rebooting of computers thus allowing the jobs to complete? In this work we relax some of the common computer management policies used in large organisations. We evaluate if alternative policies which may have impact on the primary users of the computers could save enough energy to make this impact tolerable. We evaluate our approach through the use of the HTC-Sim simulation framework. We demonstrate a potential energy saving of 12.4%, along with overhead reductions of 20--74%, and up to 48% reduction in job terminations by re-directing only 13% of users away from computers servicing HTC jobs.