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:20260522T150119Z
LOCATION:C2/3/4 Ballroom
DTSTART;TZID=America/Chicago:20181113T083000
DTEND;TZID=America/Chicago:20181113T170000
UID:submissions.supercomputing.org_SC18_sess325_spost119@linklings.com
SUMMARY:Designing Shared Address Space MPI Libraries in Many-Core Era
DESCRIPTION:Jahanzeb Maqbool Hashmi (Ohio State University)\n\nThe emergen
 ce of modern multi-/many-cores has put more emphasis on optimizing intra-n
 ode communication. Existing designs in MPI libraries that work on the conc
 ept of distributed address spaces incur the overhead of intermediate memor
 y copies to stage the data between processes. This can lead to severe perf
 ormance degradation especially on emerging many-core architectures like In
 tel Skylake and IBM OpenPOWER. This work proposes a high-performance "shar
 ed address-space"-based MPI point-to-point and collective communication de
 signs using XPMEM. We first characterize the bottlenecks associated with X
 PMEM based communication and propose new designs for efficient MPI large m
 essage communication. Then we propose novel collective designs that are co
 ntention-free and offer true zero-copy reduction operations. The proposed 
 designs are evaluated on different multi-/many-core architectures using va
 rious micro-benchmarks and application kernels such as MiniAMR and AlexNet
  DNN training on CNTK. The proposed designs have shown significant perform
 ance improvement over state-of-the-art available in MPI libraries.\n\nRegi
 stration Category: Tech Program Reg Pass, Exhibits Reg Pass\n\n
END:VEVENT
END:VCALENDAR