Using Model-based Clustering to Improve Predictions for Queueing Delay on Parallel Machines

Abstract

Most space-sharing parallel computers presently operated by production high-performance computing centers use batch-queuing systems to manage processor allocation. In many cases, users wishing to use these batch-queued resources may choose among different queues (charging different amounts) potentially on a number of machines to which they have access. In such a situation, the amount of time a user's job will wait in any one batch queue can be a significant portion of the overall time from job submission to job completion. It thus becomes desirable to provide a prediction for the amount of time a given job can expect to wait in the queue. Further, it is natural to expect that attributes of an incoming job, specifically the number of processors requested and the amount of time requested, might impact that job's wait time. In this work, we explore the possibility of generating accurate predictions by automatically grouping jobs having similar attributes using model-based clustering. Moreover, we implement this clustering technique for a time series of jobs so that predictions of future wait times can be generated in real time. Using trace-based simulation on data from 7 machines over a 9-year period from across the country, comprising over one million job records, we show that clustering either by requested time, requested number of processors, or the product of the two generally produces more accurate predictions than earlier, more naive, approaches and that automatic clustering outperforms administrator-determined clustering.