BEGIN:VCALENDAR
VERSION:2.0
PRODID:Linklings LLC
BEGIN:VTIMEZONE
TZID:America/Denver
X-LIC-LOCATION:America/Denver
BEGIN:DAYLIGHT
TZOFFSETFROM:-0700
TZOFFSETTO:-0600
TZNAME:MDT
DTSTART:19700308T020000
RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=2SU
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:-0600
TZOFFSETTO:-0700
TZNAME:MST
DTSTART:19701101T020000
RRULE:FREQ=YEARLY;BYMONTH=11;BYDAY=1SU
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTAMP:20200129T163600Z
LOCATION:607
DTSTART;TZID=America/Denver:20191118T165000
DTEND;TZID=America/Denver:20191118T171000
UID:submissions.supercomputing.org_SC19_sess124_ws_lasalss113@linklings.co
m
SUMMARY:Making Speculative Scheduling Robust to Incomplete Data
DESCRIPTION:Workshop\n\nMaking Speculative Scheduling Robust to Incomplete
Data\n\nGainaru, Pallez\n\nWe study in this work the robustness of Specul
ative Scheduling to the\nincompleteness of data. Speculative scheduling ha
s been introduce as a solution\nto incorporate future types of application
s into the design of HPC schedulers, specifically\napplications whose runt
ime is not perfectly known but can be modeled with\nprobability distributi
ons. Preliminary studies show the importance of speculative scheduling \nw
hen dealing with stochastic applications when the application runtime mode
l\nis completely known. In this work we show how one can extract even from
incomplete data on\nthe behavior of HPC applications enough information s
o that speculative scheduling performs well.\n\nSpecifically, we show that
for synthetic runtimes who follow usual probability\ndistributions such a
s truncated normal distribution, we can extract enough data\nfrom as littl
e as 10 previous runs, to be within 5\% of the solution\nwhich has all the
exact information. For real traces of applications, the\nperformance with
10 data points varies with the applications (within 20\% of the\nfull-kno
wledge solution), but converges fast (5\% with 100 previous samples).\n\nF
inally a side effect of this study is to show the importance of the theore
tical\nresults obtained on continuous probability distributions for specul
ative\nscheduling. Indeed, we observe that the solutions for such distribu
tions are\nmore robust to incomplete data than the solutions for discrete
distributions.\n\nTag: Workshop Reg Pass, Algorithms, Scalable Computing\n
\nRegistration Category: Workshop Reg Pass, Algorithms, Scalable Computing
URL:https://sc19.supercomputing.org/presentation/?id=ws_lasalss113&sess=se
ss124
END:VEVENT
END:VCALENDAR