Report ID
2005-09
Report Authors
John Brevik, Dan Nurmi, and Rich Wolski
Report Date
Abstract
Most space-sharing resources presently operated by high performancecomputing centers employ some sort of batch queueing system to manageresource allocation to multiple users.In this work, weexplore a new method for providing end-users with predictions of thebounds on queuing delay individual jobs will experience when waiting to bescheduled to a machine partition.We evaluate this methodusing scheduler logs that covera 9 year period from 7 large HPC centers. Our results show thatit is possible to predict delay bounds with specified confidence levels forjobs in different queues, and for jobs requesting differentranges of processor counts.
Document
2005-09.pdf269.44 KB