BEGIN:VCALENDAR
VERSION:2.0
PRODID:icalendar-ruby
CALSCALE:GREGORIAN
METHOD:PUBLISH
BEGIN:VEVENT
DTSTAMP:20261016T014354Z
UID:https://www.mpip-mainz.mpg.de/events/27030/536464
DTSTART:20210119T130000Z
DTEND:20210119T140000Z
CLASS:PUBLIC
CREATED:20210112T092115Z
DESCRIPTION:Parallel computing has developed as a central tool in scientifi
 c computing to solve large scale problems involving huge number of degrees
  of freedom\, complex geometries or coupled applications. The parallel eff
 iciency is key for estimating to which degree the computational resources 
 are used\, or whether there is still potential to speed up an application 
 by organising data or workflow in a different way across processors. To re
 duce the wall clock time of an application\, a goal might be to use as man
 y processors of a parallel architecture as possible. However\, scalability
  of a parallel application depends on a number of characteristics\, among 
 which is efficient communication\, equal distribution of work or efficient
  data layout.Many parallel applications\, especially particle or mesh base
 d algorithms like Molecular Dynamics or Lattice Boltzmann methods\, are im
 plemented by domain decomposition techniques\, where processors administra
 te certain geometrical regions of a physical system. In such cases\, unequ
 al work load in the processor network is to be expected when particles are
  not distributed homogeneously or the computation cost of particle interac
 tions is not equal in each part of the system. Also in the case where hete
 rogeneous architecture components are coupled together in a complex cluste
 r network (e.g. CPU-GPU\, different types of CPUs or different network spe
 eds) wall clock times for solving a problem with the same number of degree
 s of freedom will vary across the parallel application. For these scenario
 s the code has to decide how to redistribute the work among processes acco
 rding to a work sharing protocol or to dynamically adjust computational do
 mains\, to balance the workload.In the seminar\, I will give an introducti
 on to the problem of load balancing and discuss various methods to redistr
 ibute data or re-organise the domain decomposition to improve and optimise
  the work load and to improve parallel efficiency and scalability. As an o
 utlook I will discuss developments from the European Centre of Excellence 
 E-CAM\, where different methods have been implemented into a library\, whi
 ch can be used in community codes.\nVortragende(r): Prof. Godehard Sutmann
LAST-MODIFIED:20210114T195706Z
LOCATION:Digital see link
ORGANIZER;CN=Burkhard Duenweg:mailto:duenweg@mpip-mainz.mpg.de
SUMMARY:Dynamic Load Balancing for Parallel Particle Simulations
URL;VALUE=URI:https://www.mpip-mainz.mpg.de/events/27030/536464
END:VEVENT
END:VCALENDAR