On the feasibility of dynamic rescheduling on the Intel Distributed Computing Platform.

ABSTRACTThis paper examines the feasibility of dynamic rescheduling techniques for effectively utilizing compute resources within a data center. Our work is motivated by practical concerns of Intel Distributed Computing Platform (IDCP), an Internet-scale data center based distributed computing platform developed by Intel Corporation for massively parallel chip simulations within the company. IDCP has been operational for many years, and currently is deployed live on tens of thousands of machines that are globally distributed at various data centers. We perform an analysis of job execution traces obtained over a one year period collected from tens of thousands of IDCP machines from 20 different pools. Our analysis shows that the IDCP currently does not make full use of all the resources. Specifically, the job completion time can be severely impacted due to job suspension when high priority jobs preempt low priority jobs. We then develop dynamic job rescheduling strategies that adaptively restart jobs to available resources elsewhere, which better utilize system resources and improve completion times. Our trace-driven evaluation results show that dynamic rescheduling enables IDCP to significantly reduce system waste and completion time of suspended jobs.