BEGIN:VCALENDAR
VERSION:2.0
PRODID:Linklings LLC
BEGIN:VTIMEZONE
TZID:America/New_York
X-LIC-LOCATION:America/New_York
BEGIN:DAYLIGHT
TZOFFSETFROM:-0500
TZOFFSETTO:-0400
TZNAME:EDT
DTSTART:19700308T020000
RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=2SU
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:-0400
TZOFFSETTO:-0500
TZNAME:EST
DTSTART:19701101T020000
RRULE:FREQ=YEARLY;BYMONTH=11;BYDAY=1SU
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTAMP:20210402T160557Z
LOCATION:Track 7
DTSTART;TZID=America/New_York:20201113T143000
DTEND;TZID=America/New_York:20201113T150000
UID:submissions.supercomputing.org_SC20_sess226_ws_ross106@linklings.com
SUMMARY:Towards Generalizable Models of I/O Throughput
DESCRIPTION:Workshop\n\nTowards Generalizable Models of I/O Throughput\n\n
 Isakov, del Rosario, Madireddy, Balaprakash, Carns...\n\nMany modern HPC a
 pplications do not make good use of the limited available I/O bandwidth. D
 eveloping an understanding of the I/O subsystem is a critical first step i
 n order to better utilize an HPC system. While expert insight is indispens
 able, I/O experts are in rare supply. We seek to automate this effort by d
 eveloping and interpreting models of I/O throughput. Such interpretations 
 may be useful to both application developers who can use them to improve t
 heir codes and to facility operators who can use them to identify larger p
 roblems in an HPC system.\n\nThe application of machine learning (ML) to H
 PC system analysis has been shown to be a promising direction. The direct 
 application of ML methods to I/O throughput prediction, however, often lea
 ds to brittle models with low extrapolative power. In this work, we set ou
 t to understand the reasons why common methods underperform on this specif
 ic problem domain, and how to build models that better generalize on unsee
 n data. We show that commonly used cross-validation testing yields sets th
 at are too similar, preventing us from detecting overfitting. We propose a
  method for generating test sets that encourages training-test set separat
 ion. Next, we explore limits of I/O throughput prediction and show that we
  can estimate I/O contention noise by observing repeated runs of an applic
 ation. Then, we show that by using our new test sets, we can better discri
 minate different architectures of ML models in terms of how well they gene
 ralize.\n\nTag: System Software and Runtime Systems\n\nRegistration Catego
 ry: Workshop Reg Pass
END:VEVENT
END:VCALENDAR