[Paraview] Parallel Reading and Extent Translator

[Jean M. Favre]

Back in March 2009, I answered a question by another user
(http://www.paraview.org/pipermail/paraview/2009-March/011303.html),
asking how to override the default splitting of data chunks. By default,
structured volumes are split among all 3 main directions. My answer,
validated with version 3.4, was to tell the Extent Translator to use an
alternate SplitMode, like this:

vtkStreamingDemandDrivenPipeline *exec =
static_cast<vtkStreamingDemandDrivenPipeline*>(this->GetExecutive());
exec->GetExtentTranslator(outInfo)->SetSplitModeToXSlab();

I have verified today that the solution works well, producing the
required split, either as X, Y or Z slabs.

My problem is that I cannot reproduce this in version 3.6.1. When the
same data reader source code is compiled against version 3.6, my request
to SetSplitModeTo?Slab() are ignored all the times. I am using the same
grid resolution and the same number of processes. What works in version
3.4 cannot be reproduced in version 3.6.

Has anyone seen this too?

If anyone wonders, it does matter greatly. Reading HDF5 hyperslabs can
be 10 times faster in one direction or the other. That is what I need to
support.

Jean--
Swiss National Supercomputing Center.