BEGIN:VCALENDAR
VERSION:2.0
PRODID:Linklings LLC
BEGIN:VTIMEZONE
TZID:America/Chicago
X-LIC-LOCATION:America/Chicago
BEGIN:DAYLIGHT
TZOFFSETFROM:-0600
TZOFFSETTO:-0500
TZNAME:CDT
DTSTART:19700308T020000
RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=2SU
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:-0500
TZOFFSETTO:-0600
TZNAME:CST
DTSTART:19701101T020000
RRULE:FREQ=YEARLY;BYMONTH=11;BYDAY=1SU
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTAMP:20260522T150120Z
LOCATION:D220
DTSTART;TZID=America/Chicago:20181111T141500
DTEND;TZID=America/Chicago:20181111T141800
UID:submissions.supercomputing.org_SC18_sess160_ws_whpc126@linklings.com
SUMMARY:Study of Performance Variability on Dragonfly Systems
DESCRIPTION:Xin Wang (Illinois Institute of Technology)\n\nDragonfly netwo
 rks are being widely adopted in high-performance computing systems. On the
 se networks, however, interference caused by resource sharing can lead to 
 significant network congestion and performance variability. On a shared ne
 twork, different job placement policies lead to different traffic distribu
 tions. Contiguous job placement policy achieves localized communication by
  assigning adjacent compute nodes to the same job. Random job placement po
 licy, on the other hand, achieves balanced network traffic by placing appl
 ication processes sparsely across the network to uniformly distribute the 
 message load. Localized communication and balanced network traffic have op
 posite advantages and drawbacks. Localizing communication reduces the numb
 er of hops for message transfers at the cost of potential network congesti
 on, while balancing network traffic reduces potential local congestion at 
 the cost of increased message transfer hops.\n\nIn this study, we first pr
 esent a comparative analysis exploring the trade-off between localizing co
 mmunication and balancing network traffic using trace-based simulations, a
 nd demonstrate the effect of external network interference by introducing 
 background traffic and show that localized communication can help reduce t
 he application performance variation caused by network sharing. We then in
 troduce an online simulation framework that improves performance and scala
 bility, and discuss the validation of the simulation observations to a pro
 duction Dragonfly system in respect of performance variability.\n\nTag: Di
 versity, Education, Hot Topics\n\nRegistration Category: Workshop Reg Pass
 \n\nSession Chairs: Toni Collis (Women in High Performance Computing); Wer
 onika Filinger (Edinburgh Parallel Computing Centre (EPCC); University of 
 Edinburgh, Scotland); and Misbah Mubarak (Amazon Web Services)\n\n
END:VEVENT
END:VCALENDAR