Keywords: high performance computing, massively parallel, adaptive meshing, multigrid solver, Tier0, PRACE.

In this paper, work performed to allow full using Tier0 supercomputers for finite element simulations that include unstructured mesh adaptation is presented. Full use means using almost all the cores present on the system, but also all the memory available. This work has been done within the PRACE project Cim128Ki, during PRACE's 6th regular call, providing us access to two top European supercomputers: Curie and JuQUEEN. The main goal of this project was to test the scalability of our finite element mesh adaptation application on full Tier0 supercomputers (with size up to105cores). The simulations concerned are performed using an implicit finite element method built on adaptive unstructured meshes. Thus, computational time is mainly spent on two steps: resolution of a linear system; and mesh adaptation. Firstly, some definitions about massively parallel computations and their limits for finite element applications are explained. Secondly, the strategy and developments performed to adapt parallel mesh adaptation and linear solvers to enable executing them using almost all the cores and memory present on Tier0 supercomputers are detailed. Before presenting an example of simulation done using the power given by the referred supercomputers, parallel performances obtained on Tier0 supercomputers will be given, including the largest computation reached: generation of a 33.4 billion-node mesh and resolution of a linear system of 100 billion unknowns, consuming 200TB of memory, over 65,536 Intel Xeon cores on Curie and 261,144 PowerPC cores on JuQUEEN.

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)