Performance and Extension of a Particle Transport Code using Hybrid MPI/OpenMP Programming Models

Gavin Pringle, Dave Barrett, David Turland, Michele Weiland, Mark Parsons

Research output: Contribution to conferencePaperpeer-review

Abstract

We describe AWE's HPC benchmark particle transport code, which employs a wavefront sweep algorithm. After almost 4 years collaboration between EPCC and AWE, we present Chimaera-2_3D: a Fortran90 and MPI/OpenMP code which scales well to thousands of cores for large problem sizes. Significant restructuring has increased the degrees of parallelism available to efficiently exploit future many-core exascale systems. For OpenMP, we have introduced slices through the cuboid mesh which present a set of cells which may be computed independently; and computation over the angles within each cell can also be parallelized using OpenMP. Previously, the initial form of Chimaera computed a coupled, inter-dependent iteration over 'Energy Groups'. Our new code now decouples these iterations which, whilst increasing the computational time, permits a new task level of efficient parallelism encoded using MPI. This paper will present results from the extensive benchmarking exercise using a Cray XT4/5 (HECToR) and a Cray XC30 (ARCHER).Wa
Original languageEnglish
Publication statusPublished - 2015
EventCray User Group 2015 - National Centre for Supercomputing Applications, Chicago, United States
Duration: 26 Apr 201530 Apr 2015

Conference

ConferenceCray User Group 2015
CountryUnited States
CityChicago
Period26/04/1530/04/15

Keywords

  • Wavefront sweep
  • MPI
  • OpenMP
  • mixed-mode
  • Cray

Fingerprint Dive into the research topics of 'Performance and Extension of a Particle Transport Code using Hybrid MPI/OpenMP Programming Models'. Together they form a unique fingerprint.

Cite this