BEGIN:VCALENDAR
VERSION:2.0
PRODID:Linklings LLC
BEGIN:VTIMEZONE
TZID:America/New_York
X-LIC-LOCATION:America/New_York
BEGIN:DAYLIGHT
TZOFFSETFROM:-0500
TZOFFSETTO:-0400
TZNAME:EDT
DTSTART:19700308T020000
RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=2SU
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:-0400
TZOFFSETTO:-0500
TZNAME:EST
DTSTART:19701101T020000
RRULE:FREQ=YEARLY;BYMONTH=11;BYDAY=1SU
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTAMP:20260422T143141Z
LOCATION:B308
DTSTART;TZID=America/New_York:20241119T140000
DTEND;TZID=America/New_York:20241119T143000
UID:submissions.supercomputing.org_SC24_sess391_pap278@linklings.com
SUMMARY:PAL: A Variability-Aware Policy for Scheduling ML Workloads in GPU
  Clusters
DESCRIPTION:Rutwik Jain, Brandon Tran, Keting Chen, Matthew Sinclair, and 
 Shivaram Venkataraman (University of Wisconsin-Madison)\n\nLarge-scale com
 puting systems are increasingly using GPUs to enable peta- and exa-scale l
 evels of compute to meet the needs of modern applications. Given the wides
 pread and growing use of ML, including in scientific applications, optimiz
 ing clusters for ML workloads is important. However, recent work has demon
 strated that accelerators in these clusters can suffer from performance va
 riability, leading to resource under-utilization and load imbalance. In th
 is work we focus on how clusters schedulers can embrace performance variab
 ility to mitigate its effects. We design a novel cluster scheduler, PAL, w
 hich uses application-specific variability profiles to improve job perform
 ance and resource utilization. PAL also balances performance variability w
 ith locality. Overall, PAL significantly improves GPU-rich cluster schedul
 ing: across traces for six ML workloads with a variety of variability prof
 iles, PAL improves geomean job completion time by 42% and cluster utilizat
 ion by 28% over existing state-of-the-art schedulers.\n\nTag: Distributed 
 Computing, Middleware and System Software, Parallel Programming Methods, M
 odels, Languages and Environments, Programming Frameworks and System Softw
 are, Resource Management\n\nRegistration Category: Tech Program Reg Pass\n
 \nSession Chair: Dong Dai (University of Delaware)\n\n
END:VEVENT
END:VCALENDAR