I am trying to visualize data from my CFD solver using the VTKHDF file format and ParaView. The solver is MPI parallel. Until now I resorted to assembling the mesh in memory and then writing a single partition to the output file. The resulting output file worked as expected, as in when I use filters such as CellDataToPointData or Contour, I get the expected output.
For very large datasets, however, this strategy seems to not work, as I was getting some allocation failure from ParaView when opening the dataset (very big dataset >64GB). I checked the documentation again for VTKHDF and am now trying to write partitioned datasets instead. That way I can run ParaView on the distributed dataset (probably as intended).
The issue I have now is that when I visualize my data, I can see imprints of the individual partitions. I would like to get rid of these, for obvious reasons. Per rank, I am writing the points (includes duplicates from other ranks) and all inner cells (no duplicates). When I use the Contour filter, for example, the output is “closed”. What I tried was to mark the duplicated points with a vtkGhostType array and add the GlobalNodeId for all the points. Then I tried to use the AddGhostCells filter to have ParaView add ghost cells so that the contour output would be closed. This, however, crashed my ParaView. After some Googling, I found a thread that mentioned that you have to run RedistributeMesh before generating ghost cells. This, however, also does not produce the desired result.
I have attached sample data that illustrates my problem. Could someone please tell me what I am doing wrong in this case.
Steps to reproduce:
- Unzip data
- Load navier_stokes_00011.vtkhdf in ParaView
- Contour filter for density (Value=2)
- Output is not smooth but has imprints of partitions
data.zip (4.0 MB)