Repository navigation
Remove MPI from openPMD output, extract the driver’s diagnostics, refactor the collapse demos - #98
Merged
Merged
Conversation
…each process during multinode simulations
…mental_gauss_law_lapse_tracking
…g with readability
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Per-process openPMD output + offline merge (replaces MPI collective I/O)
Multiprocess runs no longer need MPI-enabled openPMD or parallel HDF5. Each JAX process writes its own data/fields_pNNN.h5 / particles_pNNN.h5 through the same bounded async queues a single-process run uses, exchanging nothing. A single-process run keeps the unsuffixed names and needs no merge.
New PyPIC3D/diagnostics/merge_openpmd.py joins a _pNNN group back into the unsuffixed file every analysis path already reads. Exposed as python -m PyPIC3D.diagnostics.merge_openpmd outputs/data and the console script PyPIC3D-merge-openpmd (--output, --iterations, --overwrite, --no-pmd). It validates that process indices are exactly 0..N-1 and that tile coverage is complete, and rewrites the .pmd pointer beside the merged series.
Because the HDF5 backend reports every dataset as one whole chunk, a reader can’t discover which regions a file wrote. The writer now records ownership explicitly as iteration attributes (pypic3dOwnedTiles, pypic3dTileCounts, pypic3dTileShape, pypic3dGlobalShape, pypic3dProcessIndex/Count), and the merge reads them back.
Mesh records use tile-shaped HDF5 chunks so a per-process file allocates only the cells that process owns, while each file stays a valid whole-domain mesh.
mpi_communicator() and every comm plumbing path are gone. initialize_distributed() now relies on JAX’s own Slurm/Open-MPI detection, with PYPIC_COORDINATOR_ADDRESS / PYPIC_PROCESS_ID / PYPIC_NUM_PROCESSES for launches it can’t detect. The [distributed] extra (mpi4py) is dropped from pyproject.toml.
PyPIC3D/distributed.py → PyPIC3D/utilities/distributed.py.
parallel_openpmd.py deleted; multihost_io_worker.py and sharded_openpmd_worker.py tests deleted, replaced by tests/distributed_tests/merge_openpmd_test.py (312 lines) which runs the whole naming/ownership/merge round trip on one process with no launcher.
docs/distributed_cartesian.md loses the parallel-HDF5/openPMD-from-source build instructions; docs/usage.rst documents the merge step.
Driver loop extraction
main.py drops from ~400 to ~210 lines. The energy/constraint/openPMD blocks move into a new PyPIC3D/diagnostics/simulation_output.py (open_simulation_output context manager, write_simulation_output, write_scalar_diagnostics, write_work_energy_tracker, simulation_energy_components, print_energy). Writer cleanup is now a finally block that closes both writers and preserves the original exception.
Supporting moves: compute_energy_components (diagnostic_quantities.py) folds the external-field add and the optional Proca term into one call; print_bssn_constraints (static_metric.py); _dark_timestep → dark_photon_courant_condition in simulation_helpers.py. Unused benchmark static parameter removed. New tests/code_tests/simulation_output_test.py.
Collapse demo refactors
Charged stellar collapse: the separate 1D radial metric solve is gone. initial_metric.py (484 lines changed, mostly deletions) is now a Cartesian time-symmetric Hamiltonian solve on the BSSN grid — JAX_BSSN’s discrete Hamiltonian, Newton with autodiff JVPs and GMRES (so memory scales with the grid rather than a dense 3D Jacobian), Robin outer boundary, backtracking line search. It is called from initial_data.py after particles and the spherical E field are loaded, so the demo is now a single python initial_data.py step.
The star is specified as a uniform coordinate sphere (rest mass 1, coordinate radius 10, Q = 0.2) with ADM mass and areal radius as measured outputs instead of prescribed inputs. charged_star.toml species values updated accordingly.
Neutral OS collapse: initial_metric.py simplified to the analytic solution, with the Wu & Most (2026) reference cited.
Readability pass
Single-call-site private helpers inlined across supergaussian.py, openPMD.py, initialization.py, particle_initialization.py, particle_tile_communication.py, grids.py, parameters.py. Notably _tile_grid_axis loses a nested lax.cond ladder that was selecting dx/dy/dz by traced axis index. Behavior-preserving; test coverage added in initialization_test.py, parameters_test.py, sharded_openpmd_test.py, distributed_particle_refresh_test.py.