Hi all,
I’ve been running some strong scaling tests with our CFD code, which has in-situ functionality using ParaView Catalyst V1. The data, we pass to the co-processor is a vtkMultiBlockDataset where only one vtkUnstructuredGrid is populated per MPI process.
I have a simple pipeline, which just writes this dataset as a VTM file. When I scale to a large number of processes (> 16k), the memory used per process when the pipeline is executed grows with the number of processes until I run out of memory on my compute nodes, although the problem size remains the same.
This does not happen if I use a pipeline where I use a MergeBlocks filter on the dataset and write as PVTU.
In my opinion, this seems as if there is a problem with the VTM writer in ParaView 5.11.2 and 5.12.1. I couldn’t test 5.13.1, yet.
Might be related to https://gitlab.kitware.com/paraview/paraview/-/issues/22671.