Latest revision as of 08:43, 1 July 2020

Abstract

This paper summarizes the major improvements and developments that have taken place during the last year for FEFLO, a general-purpose CFD code based on adaptive, unstructured grids. All aspects of a comprehensive simulation pipeline: pre-processing, gridding, field solvers and post-processing saw important advances, and are treated. The advent of machines with millions of cores focused a lot of the developments on distributed memory aspects of field solvers and multiphysics modules. Parallel grid generation, particles and flow, parallel interpolation, parallel domain splitting and repartitioning for multiphysics were all considered. Furthermore, timings on Xeon and AMD chips led to the realization that memory transfers play a considerable role as compared to floating point operations, something previous chip generations did not exhibit. This led to a thorough analysis of numerical algorithms with subsequent re-writing. This in-depth analysis determined that at present achievable speeds are limited by the communication between processors (MPI). Even as the number of cores/domains can increase to several hundred thousand, due to communication overheads the wall clock time it takes to update a flowfield is bounded. We expect that as more users get access to hundreds of thousands of cores and the domain size per core starts shrinking, they will also encounter this 'minimum timestep barrier'.