BEGIN:VCALENDAR
VERSION:2.0
PRODID:Linklings LLC
BEGIN:VTIMEZONE
TZID:Europe/Stockholm
X-LIC-LOCATION:Europe/Stockholm
BEGIN:DAYLIGHT
TZOFFSETFROM:+0100
TZOFFSETTO:+0200
TZNAME:CEST
DTSTART:19700308T020000
RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=-1SU
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:+0200
TZOFFSETTO:+0100
TZNAME:CET
DTSTART:19701101T020000
RRULE:FREQ=YEARLY;BYMONTH=10;BYDAY=-1SU
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTAMP:20210916T132454Z
LOCATION:Jean-Jacques Rousseau
DTSTART;TZID=Europe/Stockholm:20210709T140000
DTEND;TZID=Europe/Stockholm:20210709T143000
UID:submissions.pasc-conference.org_PASC21_sess130_msa207@linklings.com
SUMMARY:Experimenting with Containers to Achieve Performance Portability
DESCRIPTION:Minisymposium\n\nExperimenting with Containers to Achieve Perf
 ormance Portability\n\nLazzaro, Mujkanovic, Richardson\n\nContainers have 
 become popular for HPC applications, creating the possibility of packing e
 ntire scientific workflows, software, libraries, and even data, thus solvi
 ng the problem of making software run reliably when moved from one computi
 ng environment to another. Although the de-facto standard technology is Do
 cker, which is widely used in virtualized cloud environments, several othe
 r optimized technologies have been specifically developed for HPC systems,
  such as Charliecloud, Sarus, and Singularity. Despite the main portabilit
 y goal, the challenge for HPC applications is to achieve performance porta
 bility. In this respect, several approaches can be considered, such as dis
 patching of optimized libraries at run time, and just-in-time compilation.
  We report on performance comparison of the aforementioned container 
 technologies when running some representative HPC benchmarks on supercompu
 ters. We review the performance optimizations we applied to achieve perfor
 mance portability within the containers. Developed as part the Horizon 202
 0 SODALITE project, we also present how containerized AI training deployme
 nts can be optimized with graph compilers, which aim to optimize the execu
 tion of a Deep Neural Network (DNN) graph by generating optimized code for
  a target hardware/backend, thus accelerating training and deployment of D
 NN models.\n\nDomain: CS and Math, Physics
END:VEVENT
END:VCALENDAR