Computational & Technology Resources
an online resource for computational,
engineering & technology publications
Computational Science, Engineering & Technology Series
DEVELOPMENTS IN PARALLEL, DISTRIBUTED, GRID AND CLOUD COMPUTING FOR ENGINEERING
Edited by: B.H.V. Topping and P. Iványi
Challenges to be Overcome for Engineering Software to Run Efficiently on Petascale Machines
C. Moulinec1, D.R. Emerson1, Y. Fournier2 and P. Vezolle3
1STFC Daresbury Laboratory, SCD, Warrington, United Kingdom
C. Moulinec, D.R. Emerson, Y. Fournier, P. Vezolle, "Challenges to be Overcome for Engineering Software to Run Efficiently on Petascale Machines", in B.H.V. Topping and P. Iványi, (Editor), "Developments in Parallel, Distributed, Grid and Cloud Computing for Engineering", Saxe-Coburg Publications, Stirlingshire, UK, Chapter 2, pp 23-40, 2013. doi:10.4203/csets.31.2
Keywords: petascale, mesh generation, partitioning, solver, IO, computational fluid dynamics, PRACE.
This chapter presents some of the challenges and solutions for engineering software to run efficiently on petascale machines. Generating several billion cell meshes is almost impossible in serial, and the way forward to cirvumvent this problem, apart from using one of the rare parallel mesh generators is to combine mesh joining and mesh multiplication. Partitioning these large meshes should be run parallel either by geometric or graph-based partitioners. For matrix-based resolution, using a multigrid algorithm as a preconditioner and a deflated conjugate gradient as a solver should allow good performance. Postprocessing should be run parallel by the code itself to minimise IOs, but how to deal with the restarting procedure remains an open problem. Finally, the software should also take into account the architecture of the machines because nodes may have considerable memory capacity available, as well as many cores and sometimes GPUs.
purchase the full-text of this chapter (price £20)