The handling of XML IO in parallel in DOLFIN is presently inconsistent and
extremely slow for large files. Moreover, the code is clumsy. Now that we have
proper parallel IO support via HDF5 (and XDMF) and MPI-IO, I would like to
propose that XML IO be simplified to be serial only. Only XML meshes are
presently read in a memory scalable fashion. To read meshes in parallel,
supporting the below two implementations is suggested:
1. Recommended approach: Use XDMF/HDF5 format. To convert existing meshes:
- (simple) If mesh fits into memory, read XML mesh into DOLFIN and use
DOLFIN to write XDMF file.
- (general) Use new dolfin-xml-to-hdf5 Python script to convert XML files
to XDMF/HDF5. Script uses SAX XML parser, so memory limitations are not a
problem for large meshes.
2. Supported approach: DOLFIN reads XML mesh file on one process only and
distributes mesh to all processes. This will be considerably faster than the
present parallel implementation, which uses a slow SAX parser, but it is
limited to problems for which the mesh fits into memory on a process.
Most users will require no code changes but will see a speed-up. For large XML
files that don’t into memory on one process, HDF5/XDMF should really be used
because it is orders of magnitude faster for large problems. The proposed
change means that the XML SAX parser can be removed and libxml will no longer
be a DOLFIN dependency.
Any objections?
Garth
_______________________________________________
fenics mailing list
[email protected]
http://fenicsproject.org/mailman/listinfo/fenics
