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:20181221T160731Z
LOCATION:C140/142
DTSTART;TZID=America/Chicago:20181115T133000
DTEND;TZID=America/Chicago:20181115T140000
UID:submissions.supercomputing.org_SC18_sess190_pap425@linklings.com
SUMMARY:Exploring Flexible Communications for Streamlining DNN Ensemble Tr
 aining Pipelines
DESCRIPTION:Paper\nApplications, Cosmology, Data Analytics, Deep Learning,
  Machine Learning, Programming Systems, Storage, Visualization, Tech Progr
 am Reg Pass\n\nExploring Flexible Communications for Streamlining DNN Ense
 mble Training Pipelines\n\nPittman, Guan, Shen, Lim, Patton\n\nParallel tr
 aining of a Deep Neural Network (DNN) ensemble on a cluster of nodes is a 
 common practice to train multiple models in order to construct a model wit
 h a higher prediction accuracy. Existing ensemble training pipelines can p
 erform a great deal of redundant operations, resulting in unnecessary CPU 
 usage, or even poor pipeline performance.  In order to remove these redund
 ancies, we need pipelines with more communication flexibility than existin
 g DNN frameworks provide.\n\nThis project investigates a series of designs
  to improve pipeline flexibility and adaptivity, while also increasing per
 formance. We implement our designs using Tensorflow with Horovod, and test
  it using several large DNNs. Our results show that the CPU time spent dur
 ing training is reduced by 2-11X. Furthermore, our implementation can achi
 eve up to 10X speedups when CPU core limits are imposed. Our best pipeline
  also reduces the average power draw of the ensemble training process by 5
 -16%.
URL:https://sc18.supercomputing.org/presentation/?id=pap425&sess=sess190
END:VEVENT
END:VCALENDAR

