BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//CERN//INDICO//EN
BEGIN:VEVENT
SUMMARY:® I/O behavior of scientific deep learning workloads
DTSTART;VALUE=DATE-TIME:20231206T121000Z
DTEND;VALUE=DATE-TIME:20231206T123000Z
DTSTAMP;VALUE=DATE-TIME:20260814T183252Z
UID:indico-contribution-1971@events.chpc.ac.za
DESCRIPTION:Speakers: Hariharan Devarajan (Lawrence Livermore National Lab
 oratory)\nDeep learning has been shown as a successful method for various 
 tasks\, and its popularity results in numerous open-source deep learning s
 oftware tools. Deep learning has been applied to a broad spectrum of scien
 tific domains such as cosmology\, particle physics\, computer vision\, fus
 ion\, and astrophysics. Scientists have performed a great deal of work to 
 optimize the computational performance of deep learning frameworks. Howeve
 r\, the same cannot be said for I/O performance. As deep learning algorith
 ms rely on big-data volume and variety to effectively train neural network
 s accurately\, I/O is a significant bottleneck on large-scale distributed 
 deep learning training.\n \nIn this talk\, I aim to provide a detailed inv
 estigation of the I/O behavior of various scientific deep learning workloa
 ds running on the Theta cluster at Argonne Leadership Computing Facility. 
 In this talk\, I present DLIO\, a novel representative benchmark suite bui
 lt based on the I/O profiling of the selected workloads. DLIO can be utili
 zed to accurately emulate the I/O behavior of modern scientific deep learn
 ing applications. Using DLIO\, application developers and system software 
 solution architects can identify potential I/O bottlenecks in their applic
 ations and guide optimizations to boost the I/O performance leading to low
 er training times by up to 6.7x.\n\nhttps://events.chpc.ac.za/event/125/co
 ntributions/1971/
LOCATION:Skukuza 1-1-2 - Ndau
URL:https://events.chpc.ac.za/event/125/contributions/1971/
END:VEVENT
END:VCALENDAR