Finite elements on clusters and supercomputers
Dr Lee Margetts (ParaFEM lead developer) and Shane Balany
Vice Chancellor Room (off Pugsley Lecture Theatre Foyer) Second Floor, Queens Building.
Title: Finite elements on clusters and supercomputers
A double research talk on the use of open source finite element code ParaFEM on clusters and supercomputers. The talks will be given by Dr Lee Margetts (ParaFEM lead developer) and by Shane Balani (Mech Eng Y3 UG student).
Abstract: Dr Margetts will talk about ParaFEM [1,2], a free open source FE library that can be used to solve various types of problems, including stress analysis, heat flow, eigenvalue and forced vibrations. It has been ported to many HPC systems, including the BlueCrystal,
the UK national supercomputer ARCHER and the European level HPC systems. ParaFEM has been used to simulate dinosaur tracks, to predict heat flow in fusion reactor componets, for large scale elasto-plastic deformation and to predict crack propagation in polycrystals, among other problems.
Dr Margetts will describe the ideas behind the design of ParaFEM, it's implementation in modern Fortran and MPI. He will show examples of specific applications and performance data from varioius HPC systems.
Dr Margetts has a background in geology and geomechanics. In 2002 he obtained his PhD from the University of Manchester with the thesis "Parallel finite element analysis". He is the author and the lead developer of ParaFEM [1,2], a free open source parallel finite element library, specifically designed for clusters and supercomputers. There is a growing community of researchers in solid mechanics who are adopting ParaFEM for their work.
 I. M. Smith, D. V. Griffiths, L. Margetts, Programming the finite element method, Wiley, 5ed, 2014.
Title: Evaluating Open Source Parallel Finite Element Software as an Alternative to the Abaqus
Abstract: In this work ParaFEM was ported to a Raspberri Pi cluster, achieving a good performance per watt, that would place this cluster within the Green 500 list. ParaFEM consistently runs faster than the Abaqus for a set of studied problems. Scaling results reveal that
ParaFEM is better suited than the Abaqus for massively parallel systems. For linear elastic problems the ParaFEM results agree exactly with those obtained with the Abaqus. Discrepancies between the two codes arise in elasto-plastic analysis. These can be traced to the different solution algorithms used in the two codes. The discrepancy can be minimised with a careful selection of convergence criteria, with the use of matrix preconditioning and with a careful choice of element sizes.
Shane Balani is a 3rd year undergraduate student with the Department of Mechical Engineering. The talk summarises the findings from his individual research project.
Dr Anton Shterenlikht, Department of Mechanical Engineering