Computational & Technology Resources
an online resource for computational,
engineering & technology publications
Civil-Comp Proceedings
ISSN 1759-3433
CCP: 107
Edited by:
Paper 34

Using Full Hybrid Mesh Multiplication for Simulations on One Million MPI Tasks

A. Ronovsky1, V. Vondrak1 and C. Moulinec2

1IT4Innovations - Centre of Excellence Project, Ostrava, Czech Republic
2STFC Daresbury Laboratory, Sci-Tech Daresbury, United Kingdom

Full Bibliographic Reference for this paper
A. Ronovsky, V. Vondrak, C. Moulinec, "Using Full Hybrid Mesh Multiplication for Simulations on One Million MPI Tasks", in , (Editors), "Proceedings of the Fourth International Conference on Parallel, Distributed, Grid and Cloud Computing for Engineering", Civil-Comp Press, Stirlingshire, UK, Paper 34, 2015. doi:10.4203/ccp.107.34
Keywords: computational fluid dynamics, mesh multiplication, MPI, Code_Saturne, exascale.

Computational fluid dynamics (CFD) is one of the fields which can fully utilize the capacity of existing HPC systems. There are many cases either from basic or applied research which are so complex that their numerical simulation requires very fine representation of the computational domain. To tackle some large multi-physics multi-scale problems meshes consisting of hundred billions of cells are necessary. There are several approaches to create such huge meshes. One of them is based on global mesh refinement and is also known as mesh multiplication. Global refinement was already implemented into Code_Saturne enhancing its capability in terms of mesh refinement. Meshes with sizes of up to one hundred billion of cells were generated on-the-fly. Since there are many CFD problems where only local area is of interest (either areas close to boundaries, small geometrical entities or in regions with high gradient of solved quantities), local refinement is another approach for mesh generation. In this paper implementation of a parallel local refinement applied to Code_Saturne is described. The bottleneck of local adaptive refinement is that it breaks load balancing of the original mesh and requires a lot of global communications. Strategy to repartition the mesh before its refinement is a key issue for optimal resource utilization. To minimize the amount of data transferred among the processor cores it is necessary to do most of the communication during the preprocessing step on the original mesh before refinement. Local mesh refinement strategy was tested and its scalability and performance within Code_Saturne were analysed. Results are presented in this paper.

purchase the full-text of this paper (price £20)

go to the previous paper
go to the next paper
return to the table of contents
return to the book description
purchase this book (price £45 +P&P)