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:20260522T150116Z
LOCATION:D165
DTSTART;TZID=America/Chicago:20181112T113000
DTEND;TZID=America/Chicago:20181112T115000
UID:submissions.supercomputing.org_SC18_sess161_ws_pmbss103@linklings.com
SUMMARY:Algorithm Selection of MPI Collectives Using Machine Learning Tech
 niques
DESCRIPTION:Sascha Hunold (Technical University Wien) and Alexandra Carpen
 -Amarie (Fraunhofer Institute for Industrial Mathematics)\n\nAutotuning is
  a well established method to improve software performance for a given sys
 tem, and it is especially important in High Performance Computing. The goa
 l of autotuning is to find the best possible algorithm and its best parame
 ter settings for a given instance. Autotuning can also be applied to MPI l
 ibraries, such as OpenMPI or IntelMPI. These MPI libraries provide numerou
 s parameters that allow users to adapt them to a given system. Some of the
 se tunable parameters enable users to select a specific algorithm that sho
 uld be used internally by an MPI collective operation. For the purpose of 
 automatically tuning MPI collectives on a given system, the Intel MPI libr
 ary is shipped with mpitune. The drawback of tools like mpitune is that re
 sults can only be applied to cases (e.g., number of processes, message siz
 e) for which the tool has performed the optimization.\n\nTo overcome this 
 limitation, we present a first step towards tuning MPI libraries also for 
 unseen instances by applying machine learning techniques. Our goal is to c
 reate a classifier that takes the collective operation, the message size a
 nd communicator characteristics (number of compute nodes, number of proces
 ses per node) as an input and gives the predicted best algorithm for this 
 problem as an output. We show how such a model can be constructed and what
  pitfalls should be avoided. We demonstrate by thorough experimentation th
 at our proposed prediction model is able to outperform the default configu
 ration\n\nTag: Benchmarks, Parallel Programming Languages, Libraries, and 
 Models, Performance, Simulation\n\nRegistration Category: Workshop Reg Pas
 s\n\nSession Chair: Steven A. Wright (University of York, England)\n\n
END:VEVENT
END:VCALENDAR