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:20211207T055342Z
LOCATION:227
DTSTART;TZID=America/Chicago:20211115T143000
DTEND;TZID=America/Chicago:20211115T150000
UID:submissions.supercomputing.org_SC21_sess347_ws_scsc101@linklings.com
SUMMARY:Erasure-Coding-Based Fault Tolerance for Recommendation Model Trai
 ning
DESCRIPTION:Workshop\n\nErasure-Coding-Based Fault Tolerance for Recommend
 ation Model Training\n\nLiu, Kosaian, Vinayak\n\nDeep-learning-based recom
 mendation models (DLRMs) are widely deployed to serve personalized content
 . DLRMs are trained by distributing the model across the memory of tens/hu
 ndreds of servers. Server failures are common in such settings, and must b
 e mitigated for training to progress. Checkpointing is the primary approac
 h used for fault tolerance in these systems, but incurs significant traini
 ng-time overhead. As checkpointing overhead increases with DLRM size, chec
 kpointing is slated to become an even larger overhead as DLRMs grow. This 
 calls for rethinking fault-tolerant DLRM training.\n\nWe present ECRM, a D
 LRM training system that achieves efficient fault tolerance using erasure 
 coding. ECRM chooses which DLRM parameters to encode, efficiently updates 
 parities, and enables training to proceed without pauses. Compared to chec
 kpointing, ECRM  reduces training-time overhead, recovers from failures fa
 ster, and allows training to proceed during recovery. These results show t
 he promise of erasure coding in enabling efficient fault tolerance for DLR
 M training.\n\nTag: Parallel Programming Languages and Models, Reliability
  and Resiliency\n\nRegistration Category: Workshop Reg Pass
END:VEVENT
END:VCALENDAR