- First parallel implementation of HDF5 writer. (activated when using XDMF as output format)
- synchronous: computation->IO->computation->IO ..
- The collective MPI IO is used as middleware.
- Each process is writing its own data. This includes all ghost cells and nodes which can be viewed as a multi-block composite via xdmf. The inclusion of ghost cells will enable a simulation restart at already computer time step (not implemented yet).
- The writing algorithm is not tuned for performance yet. Performance can be compared with VTK option, with this restriction: either use real world data or disable compression for both, see explanation in !3292 (merged)
- all partition are written into a single hyperslab. Compression is off
- views to HDF5 file via XDMF file. Default xdmf generated is view to first block (first partition)
- parallel writing algorithm supports all features introduced for serial writing in !3246 (merged)
-
Feature description was added to the changelog -
Tests covering your feature were added? -
Any new feature or behavior change was documented?
Next steps:
- multi-write (for each attribute) to single-write
- alternative to hyperslab to allow compression for each partition
- analysis of bottleneck (darshan)
- chunk-size
- one process per node is writing data
- async io