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:20211207T054810Z
LOCATION:230-231-232
DTSTART;TZID=America/Chicago:20211118T110000
DTEND;TZID=America/Chicago:20211118T113000
UID:submissions.supercomputing.org_SC21_sess180_pap596@linklings.com
SUMMARY:Efficient Scaling of Dynamic Graph Neural Networks
DESCRIPTION:Paper\n\nEfficient Scaling of Dynamic Graph Neural Networks\n\
 nChakaravarthy, Pandian, Raje, Sabharwal, Suzumura...\n\nWe present distri
 buted algorithms for training dynamic Graph Neural Networks (GNN) on large
  scale graphs spanning multi-node, multi-GPU systems. To the best of our k
 nowledge, this is the first scaling study on dynamic GNN. We devise mechan
 isms for reducing the GPU memory usage and identify two execution time bot
 tlenecks: CPU-GPU data transfer; and communication volume. Exploiting prop
 erties of dynamic graphs, we design a graph difference-based strategy to s
 ignificantly reduce the transfer time. We develop a simple, but effective 
 data distribution technique under which the communication volume remains f
 ixed and linear in the input size, for any number of GPUs. Our experiments
  using billion-size graphs on a system of 128 GPUs shows that: (i) the dis
 tribution scheme achieves up to 30x speedup on 128 GPUs; (ii) the graph-di
 fference technique reduces the transfer time by a factor of up to 4.1x and
  the overall execution time by up to 40%.\n\nTag: Machine Learning and Art
 ificial Intelligence\n\nRegistration Category: Tech Program Reg Pass
END:VEVENT
END:VCALENDAR