Queueing Solutions for Scheduling Today’s Data Center Jobs
Most queueing models assume that each job runs on a single server. But this one-server-per-job model is not a good representation of today’s compute jobs, particularly Machine Learning jobs.
A typical data center job today occupies multiple cores concurrently, often thousands of cores. We refer to a job that concurrently occupies multiple cores as a multiserver job. Unfortunately, very little is known about response time in multiserver job queueing models. We present the first results on minimizing response time for multiserver job queueing models.
We also consider today’s parallel speedup jobs, which can run on any number of cores, but whose speed depends on the number of cores on which the job is run. Here it is even more complicated to understand how to best share a limited number of cores among a stream of jobs, each governed by a different speedup function. We discuss some recent optimality results in this nascent area.
Speaker: Mor Harchol-Balter, Carnegie Mellon University
Room 3108
Monday, 02/10/25
Contact:
Website: Click to VisitCost:
FreeSave this Event:
iCalendarGoogle Calendar
Yahoo! Calendar
Windows Live Calendar