BEGIN:VCALENDAR
VERSION:2.0
PRODID:Linklings LLC
BEGIN:VTIMEZONE
TZID:America/Denver
X-LIC-LOCATION:America/Denver
BEGIN:DAYLIGHT
TZOFFSETFROM:-0700
TZOFFSETTO:-0600
TZNAME:MDT
DTSTART:19700308T020000
RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=2SU
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:-0600
TZOFFSETTO:-0700
TZNAME:MST
DTSTART:19701101T020000
RRULE:FREQ=YEARLY;BYMONTH=11;BYDAY=1SU
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTAMP:20260422T000713Z
LOCATION:704-706
DTSTART;TZID=America/Denver:20231112T114200
DTEND;TZID=America/Denver:20231112T120600
UID:submissions.supercomputing.org_SC23_sess424_ws_ross106@linklings.com
SUMMARY:Fine-Grained Accelerator Partitioning for Machine Learning and Sci
 entific Computing in Function as a Service Platform
DESCRIPTION:Aditya Dhakal and Philipp Raith (Hewlett Packard Labs), Logan 
 Ward (Argonne National Laboratory (ANL)), Rolando P. Hong Enriquez and Gou
 rav Rattihalli (Hewlett Packard Labs), Kyle Chard (University of Chicago),
  Ian Foster (Argonne National Laboratory (ANL)), and Dejan Milojicic (Hewl
 ett Packard Labs)\n\nFunction-as-a-service (FaaS) is a promising execution
  environment for high-performance computing (HPC) and machine learning (ML
 ) applications, as it offers developers a simple way to write and deploy p
 rograms. Nowadays, GPUs and other accelerators are indispensable for HPC a
 nd ML workloads. However, we have observed that state-of-the-art FaaS fram
 eworks usually treat accelerators as a single device to run a single workl
 oad and have little support for multiplexing accelerators.\n\nIn this work
 , we have presented techniques to multiplex GPUs with Parsl, a popular Faa
 S framework. With our enhancements, we show up to 60% lower task completio
 n time and 250% improvement in the throughput of a large language model wh
 en multiplexing a GPU vs running without multiplexing. We plan to extend t
 he support for GPU multiplexing in FaaS platforms by tackling the challeng
 es of changing compute resources in the partition and approximating how to
  right-size a GPU partition for a function.\n\nTag: Middleware and System 
 Software, Programming Frameworks and System Software, Runtime Systems\n\nR
 egistration Category: Workshop Reg Pass\n\nSession Chairs: Balazs Gerofi (
 Intel Corporation, RIKEN Center for Computational Science (R-CCS)); Torste
 n Hoefler (ETH Zürich, Microsoft Corporation); and Kamil Iskra (Argonne Na
 tional Laboratory (ANL))\n\n
END:VEVENT
END:VCALENDAR