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:20181221T160727Z
LOCATION:D165
DTSTART;TZID=America/Chicago:20181112T110000
DTEND;TZID=America/Chicago:20181112T113000
UID:submissions.supercomputing.org_SC18_sess161_ws_pmbsf119@linklings.com
SUMMARY:Benchmarking Machine Learning Methods for Performance Modeling of 
 Scientific Applications
DESCRIPTION:Workshop\nBenchmarks, Parallel Programming Languages, Librarie
 s, and Models, Performance, Simulation, Workshop Reg Pass\n\nBenchmarking 
 Machine Learning Methods for Performance Modeling of Scientific Applicatio
 ns\n\nMalakar, Balaprakash, Vishwanat, Morozov, Kumaran\n\nPerformance mod
 eling is an important and active area of research in high-performance comp
 uting (HPC). It helps in better job scheduling and also improves overall p
 erformance of coupled applications. Sufficiently rich analytical models ar
 e challenging to develop, however, because of interactions between differe
 nt node components, network topologies, job interference, and application 
 complexity. When analytical performance models become restrictive because 
 of application dynamics and/or multicomponent interactions, machine-learni
 ng-based performance models can be helpful. While machine learning (ML) me
 thods do not require underlying system or application knowledge, they are 
 efficient in learning the unknown interactions of the application and syst
 em parameters empirically using application runs. We present a benchmark s
 tudy in which we evaluate eleven machine learning methods for modeling the
  performance of four representative scientific applications that are irreg
 ular and with skewed domain configurations on four leadership-class HPC pl
 atforms. We assess the impact of feature engineering, size of training set
 , modern hardware platforms, transfer learning, extrapolation on the predi
 ction accuracy, and training and inference times. We find that bagging, bo
 osting, and deep neural network ML methods are promising approaches with m
 edian R^2 values greater than 0.95, and these methods do not require featu
 re engineering. We demonstrate that cross-platform performance prediction 
 can be improved significantly using transfer learning with deep neural net
 works.
URL:https://sc18.supercomputing.org/presentation/?id=ws_pmbsf119&sess=sess
 161
END:VEVENT
END:VCALENDAR