BEGIN:VCALENDAR
VERSION:2.0
PRODID:Linklings LLC
BEGIN:VTIMEZONE
TZID:Europe/Stockholm
X-LIC-LOCATION:Europe/Stockholm
BEGIN:DAYLIGHT
TZOFFSETFROM:+0100
TZOFFSETTO:+0200
TZNAME:CEST
DTSTART:19700308T020000
RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=-1SU
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:+0200
TZOFFSETTO:+0100
TZNAME:CET
DTSTART:19701101T020000
RRULE:FREQ=YEARLY;BYMONTH=10;BYDAY=-1SU
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTAMP:20220812T074335Z
LOCATION:Rio Room
DTSTART;TZID=Europe/Stockholm:20220629T120000
DTEND;TZID=Europe/Stockholm:20220629T123000
UID:submissions.pasc-conference.org_PASC22_sess163_msa259@linklings.com
SUMMARY:Leveraging Heterogeneous Architectures for Molecular Dynamics at E
 xascale
DESCRIPTION:Minisymposium\n\nLeveraging Heterogeneous Architectures for Mo
 lecular Dynamics at Exascale\n\nHardy\n\nHeterogeneous architectures at ex
 ascale present significant challenge to applications like molecular dynami
 cs that are latency bound while needing to achieve strong scaling on fixed
  size problems. The NAMD parallel molecular dynamics application was from 
 inception designed for massively distributed parallelism, decomposing comp
 utation into fine-grained work objects to be executed asynchronously acros
 s hundreds of thousands of CPU cores. However, effective use of GPU accele
 rators requires aggregating these work objects into workloads of sufficien
 t size to hide the added latencies of host-device communication and GPU ke
 rnel launching. Upcoming exascale computers will feature multiple GPUs per
  node, in which CPUs provide an insignificant fraction of the peak compute
  capability. These GPU-dense nodes are capable of achieving very good sing
 le-node performance for molecular system sizes in the range of 1M to 10M a
 toms, which can be used together with multiple copy simulation techniques,
  like replica-exchange, to provide a scalable multi-node approach for acce
 ssing longer timescales. This presentation outlines the past few years of 
 NAMD development to leverage GPU-dense nodes, which has involved transitio
 ning from GPU-offload to GPU-resident approach, scaling the GPU-resident i
 mplementation across tightly coupled GPUs, and exploiting task-based paral
 lelism to load balance the non-scaling particle-mesh Ewald method for long
 -range electrostatics.\n\nDomain: Chemistry and Materials, Computer Scienc
 e and Applied Mathematics, Life Sciences, Physics
END:VEVENT
END:VCALENDAR