Talk (non-conference) (Other) FZJ-2018-04067

http://join2-wiki.gsi.de/foswiki/pub/Main/Artwork/join2_logo100x88.png
Towards Efficient Linear Scaling Green Function DFT Calculations



2017

Seminar at Center for Computational Sciences, University of Tsukuba, TsukubaTsukuba, Japan, 17 Oct 2017 - 20 Oct 20172017-10-172017-10-20

Please use a persistent id in citations:

Abstract: Materials research is an area that is expected to strongly benefit from the growing performance capabilities of future supercomputers towards Exascale. Density functional theory (DFT) has become one of the most important methods for numerical materials science. In this talk I present a methodology for restructuring a scalable DFT application for the efficient usage of current and future HPC systems based on many core architectures. Particular focus is set to Xeon Phis and GPU-accelerated systems since large HPC installations in the pre-Exascale range are planned and the compute capabilities of the first machine reaching 1 ExaFlop/s is likely to consist of one of these architectures. For the purpose of tuning and testing, a mini-app has been extracted that replays the most performance critical kernel. Inside the kernel routine, a reordering algorithm is promising to boost the performance by reducing the bandwidth requirements of a large number of small matrix-matrix multiplications.


Contributing Institute(s):
  1. Jülich Supercomputing Center (JSC)
Research Program(s):
  1. 511 - Computational Science and Mathematical Methods (POF3-511) (POF3-511)

Appears in the scientific report 2018
Database coverage:
OpenAccess
Click to display QR Code for this record

The record appears in these collections:
Document types > Presentations > Talks (non-conference)
Workflow collections > Public records
Institute Collections > JSC
Publications database
Open Access

 Record created 2018-07-09, last modified 2021-01-29


OpenAccess:
Download fulltext PDF
External link:
Download fulltextFulltext by OpenAccess repository
Rate this document:

Rate this document:
1
2
3
 
(Not yet reviewed)