[MOAB-dev] r1683 - in MOAB/trunk: . parallel

[Jason Kraftcheck]

kraftche at mcs.anl.gov wrote:
> Author: kraftche
> Date: 2008-03-24 12:21:57 -0500 (Mon, 24 Mar 2008)
> New Revision: 1683
> 
> Modified:
>    MOAB/trunk/MBWriteUtil.cpp
>    MOAB/trunk/MBWriteUtil.hpp
>    MOAB/trunk/MBWriteUtilIface.hpp
>    MOAB/trunk/RangeMap.hpp
>    MOAB/trunk/WriteHDF5.cpp
>    MOAB/trunk/WriteHDF5.hpp
>    MOAB/trunk/parallel/WriteHDF5Parallel.cpp
> Log:
> Use RangeMap rather than tag data for mapping from entity handles to file 
> IDs when writing .H5M files.  For most cases, this will result in a substantial
> reduction in the memory required to write .h5m files.  It also removes the
> limitation of 2M entities being written to a file that resuled from using
> 4-byte signed integer tag values for file IDs.*

I meant 2G (2,147,483,647), not 2M.

> 
> For a large kD-tree decomposition of a triangle mesh, write speed increased
> by about 4%.  For a large tet mesh (3.3 M tet block w/ skin triangles), 
> the write speed is 40% faster.
> 
> * Some portions of the parallel writer still have 2M limits because
>   MPI uses 'int' for sizes in calls such as MPI_Allgatherv and the
>   code does not chunk data to work around this limitation.

Here, too.

- jason