ParFlow improvements and bug-fixes would not be possible without contributions of the ParFlow community. Thank you for all the great contributions.
- ParFlow Google Group
- Default I/O mode changed to amps_sequential_io
- CLM Solar Zenith Angle Calculation
- Parallel NetCDF dataset compression
- Kokkos support
- Output of Van Genuchten variables
- Python interface updates
- Update Hypre testing to v2.18.2
- MPI runner change
- Python Interface
- Segmentation fault at end of simulation run with Van Genuchten
- Memory errors when rank contained no active cells
- PFMGOctree solver
- GFortran compilation errors
- CMake CI fixes
- CLM initialization bug
- CMake cleanup
- Fixed compilation issues in sequential amps layer
- CI has moved to Google Actions
- Sponsors acknowledgment
- PFModule extended to support output methods
- Hypre SMG and PFMG
ParFlow has switched to using a Google group for discussion from the previous email list server.
https://groups.google.com/g/parflow
Change the default I/O model to amps_sequential_io because this is the most common I/O model being used.
Add slope and aspect when determining the solar zenith angle in CLM. A new key was added Solver.CLM.UseSlopeAspect for the inclusion of slopes when determining solar zenith angles
Added configurable deflate (zlib) compression capabilities to the NetCDF based writing routines. The possibilities of parallel data compression will only work in combination with the latest NetCDF4 v4.7.4 release.
pfset NetCDF.Compression True Enable deflate based compression (default: False) pfset NetCDF.CompressionLevel 1 Compression level (0-9) (default: 1)
This work was implemented as part of EoCoE-II project (www.eocoe.eu).
Benchmark tests show that the datasize for regular NetCDF output files could be lowered by a factor of three, in addition, as less data is written to disk, the reduced data size can also lower the overall I/O footprint towards the filesystem. Therefore, depending of the selected setup, the compression overhead can be balanced by reduced writing times.
User instructions on how to use Kokkos backend can be found from README-GPU.md.
Add Kokkos accelerator backend support as an alternative to the native ParFlow CUDA backend to support more accelerator devices. The implementation does not rely on any CUDA-specific arguments but still requires Unified Memory support from the accelerator devices. It should be compatible with AMD GPUs when sufficient Unified Memory support is available.
The performance of using CUDA through the Kokkos library is slightly worse in comparison to the ParFlow native CUDA implementation. This is because a general Kokkos implementation cannot leverage certain CUDA features such as cudaMemset() for initialization or CUDA pinned host/device memory for MPI buffers. Also, Kokkos determines grid and block sizes for compute kernels differently.
The RMM pool allocator for Unified Memory can be used with Kokkos (when using Kokkos CUDA backend) and improves the performance very significantly. In the future, a Unified Memory pool allocator that supports AMD cards is likely needed to achieve good performance.
Performance of the simulation initialization phase has been improved significantly when using GPUs (with CUDA and Kokkos).
Add output for Van Genuchten values alpha, n, sres, ssat. The new output will be generated when the print_subsurf_data key is set.
The version of Hypre used for testing was updated to v2.18.2. This matches XSDK 0.5 version requirements.
The method used to find automatically find the MPI runner (mpiexec, srun etc) is based purely on the CMake FindMPI script. This should be invisible to most users.
The Beta Python interface continues to be developed. Many improvements and bugfixes have been made.
- Add hydrology functions
- CLM API bug fixes
- CLM ET calculation
- Allow clm_output function to return only 2D arrays
- Add irrigation to CLM variables
- dx/dy/dz support when writing PFB files
- Python testing support was added
- New feature to only show validation results for errors
- Table builder update: adding databases, cleanup
- Domain builder helper with examples, docs
Segmentation fault when freeing memory at the end of simulation.
The computation of the real space z vector was running beyond temporary array (zz) resulting in memory errors.
PFMGOctree was not inserting the surface coefficients correctly into the matrix with overland flow enabled.
Fixed GFortran compilation errors in ifnan.F90 with later GNU releases. Build was tested against GNU the 10.2.0 compiler suite.
On some systems, it is necessary for any binary compiled with mpi to
be executed with the appropriate
Fixed CLM bug causing long initialization times.
Updated CMake to more current usage patterns and CMake minor bugfixes.
The AMPS sequential layer had several bugs preventing it from compiling. Tests are passing again with a sequential build.
TravisCI integration for CI has been replaced with Google Actions.
A new file has been added (SPONSORS.md) to enable acknowledgment the sponsors of ParFlow development. Please feel free to submit a pull request if you wish to add a sponsor.
The testing framework has been refactored to support Python. Directory structure for tests has changed.
Add support in PFModule for module output. Two new methods were added to the PFModule 'class' to output time variant and time invariant data. This allows modules to have methods on each instance for generating output directly from the module. Previously the approach was to copy data to a problem data variable and output from the copy.
Refactored common Hypre setup code to a method to keep Hypre setup consistent.
See https://github.com/parflow/parflow/issues for current bug/issue reports.