Major changes to CarpetIOHDF5 recovery routines:

 * better general error checking and reporting
   - check return code of HDF5 API calls
   - replace asserts with CCTK_VWarns of level 0 or 1 where appropriate
   - don't just assert (0) if no valid checkpoint file was found
 * allow relaxed recovery mode (don't stop, only warn about variables which
   don't exist at recovery time)
 * MPI optimizations: gather lots of individual MPI_Bcasts into a single one
   for each datatype

Tested with CarpetIOHDF5's checkpointing/recovery testsuite.
In case of problems, it should be easy to roll back to the previous state:
the old code is still there in iohdf5chckpt_recover.cc.

Things left to do:

 * merge recovery and filereader code
 * review of checkpointing code
 * rename I/O parameters
 * reduce the set of attributes to a minimum

darcs-hash:20040707090105-1d9bf-7143ebaf80957ef45be9d9a045b5b78c27eaf8d3.gz

0 files changed, 0 insertions, 0 deletions