This section shows benchmark results using several Amazon EC2 instances for the OpenFOAM motorbike test case that is part of the OpenFOAM tutorial test-case collection. They have all been run using AWS ParallelCluster in the same way as described in this workshop however this test-case is different than the one used in this workshop, which was only 4M cells.
This case, whilst idealized, has many of the components of a more realistic external aerodynamics OpenFOAM case. It consists of a motorbike geometry in a single region with a mesh of approximately 28m hex+prism cells.
The case has been modified from the original case to increase the mesh size and also to use different domain decomposition for the meshing and solve (which provides the best HPC performance for larger cases). The timing shown in the graphs was calculated over the first 50 iterations and has then been recalculated as the number of iterations per minute to aid visualization.
All three graphs show the clear improvement from using the network optimized instance; c5n.18xlarge which includes the Elastic Fabric Adapter (EFA). The C5.24xlarge and C5a.24xlarge lose scaling earlier largely because they only have 25Gbit/s of network bandwidth. However it should be noted that the differences are only noticable at lower cells per core, although this is dependant on the problem size and OpenFOAM settings e.g solver type.
These graphs are shown per node, per core and cells per core to aid visualization of the results.