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:20210527T134945Z
LOCATION:Digital
DTSTART;TZID=Europe/Stockholm:20200623T132500
DTEND;TZID=Europe/Stockholm:20200623T135000
UID:isc_hpc_ISC High Performance 2020_sess341_pap162@linklings.com
SUMMARY:Desynchronization and Wave Pattern Formation in MPI-Parallel and H
 ybrid Memory-Bound Programs
DESCRIPTION:Research Paper\n\nDesynchronization and Wave Pattern Formation
  in MPI-Parallel and Hybrid Memory-Bound Programs\n\nAfzal, Hager, Wellein
 \n\nAnalytic, first-principles performance modeling of\ndistributed-memory
  parallel codes is notoriously imprecise. Even for\napplications with extr
 emely regular and homogeneous\ncompute-communicate phases, simply adding c
 ommunication time to\ncomputation time does often not yield a satisfactory
  prediction of\nparallel runtime due to deviations from the expected simpl
 e lockstep\npattern caused by system noise, variations in communication ti
 me,\nand inherent load imbalance. In this paper, we highlight the\nspecifi
 c cases of provoked and spontaneous desynchronization of\nmemory-bound, bu
 lk-synchronous pure MPI and hybrid MPI+OpenMP\nprograms. Using simple micr
 obenchmarks we observe that although\ndesynchronization can introduce incr
 eased waiting time per process,\nit does not necessarily cause lower resou
 rce utilization but can lead\nto an increase in available bandwidth per co
 re. In case of\nsignificant communication overhead, even natural noise can
  shove the\nsystem into a state of automatic overlap of communication and\
 ncomputation, improving the overall time to solution. The saturation\npoin
 t, i.e., the number of processes per memory domain required to\nachieve fu
 ll memory bandwidth, is pivotal in the dynamics of this\nprocess and the e
 merging stable wave pattern. We also demonstrate\nhow hybrid MPI-OpenMP pr
 ogramming can prevent desirable\ndesynchronization by eliminating the band
 width bottleneck among\nprocesses.  A Chebyshev filter diagonalization app
 lication is used\nto demonstrate some of the observed effects in a realist
 ic setting.\n\nTag: Pre-Recorded
URL:https://2020.isc-program.com/presentation/?id=pap162&sess=sess341
END:VEVENT
END:VCALENDAR

