Title:

Scheduling Multiserver Compute Jobs

Abstract:

Almost all queueing models assume that a job runs on a single server. But this one-server-per-job model is not a good representation of today's compute jobs.   A typical data center job today occupies multiple cores concurrently.  We refer to a job that requires a fixed number of cores for some duration as a multiserver job.  Unfortunately, very little is known about the performance of multiserver job queueing models.  We present the first results on response time for multiserver job models. To do this, we introduce a new scheduling policy, called ServerFilling, which is both throughput-optimal in many settings and also lends itself to response time analysis.  

 Joint work with:  Isaac Grosof and Alan Scheller-Wolf  

Bio:

Mor Harchol-Balter is the Bruce J. Nelson Professor of Computer Science at Carnegie Mellon University.  She is a Fellow of both ACM and IEEE.  She currently serves as SIG Chair for ACM SIGMETRICS, and has previously served as General Chair and TPC Chair for SIGMETRICS. She is the recipient of the NSF CAREER award, dozens of Industrial Faculty Awards, and several teaching awards, including the Herbert A. Simon Teaching Award and the Spira Teaching Award. Mor is the author of a popular queueing theory textbook, “Performance Analysis and Design of Computer Systems,” published by Cambridge University Press 2013.  She also has a new textbook coming out called “Introduction to Probability for Computing,” published by Cambridge University Press 2024. Mor’s work has been honored with many paper awards: INFORMS George Nicholson Prize 22, SIGMETRICS 21, SIGMETRICS 19, PERFORMANCE 18, INFORMS APS 18, EUROSYS 16, MASCOTS 16, MICRO 10, SIGMETRICS 03, ITC 03, SIGMETRICS 96.  She has also been blessed with fantastic PhD students, almost all of whom are professors at top universities.