[Paraview-developers] "Unable to build k-d tree structure" error with D3 filter and PV 4.4.0 on large datasets

[Schlottke-Lakemper, Michael]

Hi Joachim,

Thanks a lot, this looks like a very interesting alternative indeed! This option triggers a few more questions for me (I hope that’s ok):

- The docs indicate that the filter might perform faster if global point ids are available. Can you give a ballpark estimate about how much of a speedup we’re talking here?
- How can I make global point ids available that are suitable for this filter? Please note that we are using a custom reader plugin.
- Is it possible to generate more than one ghost layer and if yes, is this handled transparently or do I need to configure this somewhere?
- Is it possible/feasible/sensible to call this filter directly from within our custom reader plugin, such that users do not need to apply the filter manually each time?
- And finally: up to how many ranks does this filter scale (or at least has been tested and found to behave satisfactorily) - could you give me a rough estimate?

Sorry for the many questions - please let me know if you would rather have me send them in separate mails. At the moment we roll our own code for ghost cell generation, but the current implementation does not scale at all (we are ok for ~10 ranks but need to support 100-1000…) and I am looking forward to ditching it completely in favor of the filter developed by you.

Regards,

Michael

> On 25 Mar 2016, at 05:36 , Joachim Pouderoux <joachim.pouderoux at kitware.com> wrote:
> 
> Michael,
> 
> If you just intend to compute ghost cells on your unstructured grid, I would recommend you to use the new "Generate Ghost Cells" filter available in ParaView 5.0. It requires less memory than D3 and will perform faster if you don't require data redistribution.
> Having it as a plugin on PV 4.4 should be straightforward as this is a single class (vtkPUnstructuredGridGhostCellsGenerator) you will have to pick in VTK 7.0.
> 
> Best,
> Joachim