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:20210402T160556Z
LOCATION:Track 5
DTSTART;TZID=America/New_York:20201112T125800
DTEND;TZID=America/New_York:20201112T132700
UID:submissions.supercomputing.org_SC20_sess211_ws_pdsw109@linklings.com
SUMMARY:GPU Direct I/O with HDF5
DESCRIPTION:Workshop\n\nGPU Direct I/O with HDF5\n\nRavi, Byna, Koziol\n\n
 Exascale HPC systems are  being  designed  with accelerators, such as GPUs
 , to accelerate parts of applications.  In machine learning workloads as w
 ell as large-scale simulations that use GPUs as accelerators, the CPU (or 
 host) memory is currently used as a buffer for data transfers between GPU 
 (or device) memory and the file system.  If the CPU does not need to opera
 te on the data, then this is sub-optimal because it wastes host memory by 
 reserving space for duplicated data.  Furthermore, this “bounce buffer” ap
 proach wastes CPU cycles spent on transferring data.  A new technique, Nvi
 dia GPU DirectStorage (GDS), can eliminate the need to use the host memory
  as a bounce buffer.  Thereby, it becomes possible to transfer data direct
 ly between the device memory and the file system.  To take full advantage 
 of GDS in existing applications, it is necessary to provide support with e
 xisting I/O libraries, such as HDF5 and MPI-IO, which are heavily used in 
 applications.\n\nIn this paper, we describe our effort of integrating GDS 
 with HDF5, the top I/O library at NERSC and at DOE leadership computing fa
 cilities.  We design and implement this integration using a HDF5 Virtual F
 ile Driver (VFD). The GDS VFD provides a file system abstraction to the ap
 plication that allows HDF5 applications to perform I/O without the need to
  move data between CPUs and GPUs explicitly. We compare performance of the
  HDF5 GDS VFD with explicit data movement approaches and demonstrate super
 ior performance with the GDS method.\n\nTag: Big Data, Data Analytics, Com
 pression, and Management, Data Movement, File Systems and I/O, Storage\n\n
 Registration Category: Workshop Reg Pass
END:VEVENT
END:VCALENDAR