Global ops reduction

[nfarabullini]

Implementation of global max and sum for MPI operations in the same fashion as min:
Addition of global_min reduction for nflat_gradp:

    def global_min(self, buffer: data_alloc.NDArray) -> state_utils.ScalarType:
        min_buffer_arr = self._xp.array([self._xp.min(buffer)])
        recv_buffer = self._xp.empty(1, dtype=buffer.dtype)
        self.props.comm.Allreduce(min_buffer_arr, recv_buffer, mpi4py.MPI.MIN)
        return recv_buffer.item()

Note: mpi4py.MPI does not have a mean operator

[msimberg]

Two minor requests, sorry for not spotting them earlier. Let's get this in after that.

[msimberg]

Commenting for future reference: if we wanted to avoid this we could supply the neutral element as an initial value.

No change needed now, this is fine.

[nfarabullini]

cscs-ci run default

[msimberg]

Thanks, looks good to me. Let's hope the CI comes back clean afterwards (alps is down now, that's why the latest run failed).

I'll try to remove the torus xfails in #692.

[Copilot]

Pull request overview

This PR implements global reduction operations (min, max, sum, and mean) for MPI environments to enable proper distributed computation across multiple processes. The main focus is adding MPI-aware global reductions that perform both local and global (cross-process) reduction operations using mpi4py.

Changes:

• Added Reductions protocol and implementations (SingleNodeReductions and GlobalReductions) for min/max/sum/mean operations
• Updated compute_nflat_gradp to use global min reduction for proper MPI support
• Added exchange parameter to compute_coeff_gradekin for halo exchange
• Updated serialized test data URLs for multiple experiments to support 2 and 4 process configurations
• Changed do_exchange to True for inverse field computations in grid geometry

Reviewed changes

Copilot reviewed 17 out of 17 changed files in this pull request and generated 11 comments.

Show a summary per file

File	Description
model/common/src/icon4py/model/common/decomposition/definitions.py	Adds Reductions protocol and SingleNodeReductions implementation, plus factory function
model/common/src/icon4py/model/common/decomposition/mpi_decomposition.py	Implements GlobalReductions class with MPI-aware reduction operations
model/common/src/icon4py/model/common/metrics/metric_fields.py	Updates compute_nflat_gradp to use configurable min reduction
model/common/src/icon4py/model/common/metrics/metrics_factory.py	Passes global reductions to metric field computations
model/common/src/icon4py/model/common/metrics/compute_coeff_gradekin.py	Adds exchange parameter for halo synchronization
model/common/src/icon4py/model/common/grid/icon.py	Adds mean_reduction parameter to GlobalGridParams.from_fields
model/common/src/icon4py/model/common/grid/grid_manager.py	Integrates global reductions into grid manager
model/common/src/icon4py/model/common/grid/geometry.py	Enables halo exchange for inverse field computations
model/standalone_driver/src/icon4py/model/standalone_driver/standalone_driver.py	Creates and passes global reductions to grid manager
model/standalone_driver/src/icon4py/model/standalone_driver/driver_utils.py	Updates signature to accept global_reductions parameter
model/testing/src/icon4py/model/testing/definitions.py	Updates serialized data URLs for multi-process test configurations
model/testing/src/icon4py/model/testing/fixtures/datatest.py	Adds skip for GAUSS3D experiment with MPI
model/common/tests/common/decomposition/mpi_tests/test_mpi_decomposition.py	Adds comprehensive tests for all global reduction operations
model/common/tests/common/metrics/mpi_tests/test_parallel_metrics.py	Adds test for distributed nflat_gradp computation
model/common/tests/common/fixtures.py	Integrates global reductions into test fixtures
model/common/tests/common/metrics/unit_tests/test_compute_coeff_gradekin.py	Updates test to pass dummy exchange function
model/common/tests/common/grid/mpi_tests/test_parallel_geometry.py	Updates imports for test definitions

Comments suppressed due to low confidence (2)

model/common/src/icon4py/model/common/metrics/compute_coeff_gradekin.py:32

• The docstring is outdated. It mentions horizontal_end parameter which doesn't exist in the function signature. The new parameters exchange and array_ns are not documented.

    """
    Compute coefficients for improved calculation of kinetic energy gradient

    Args:
        edge_cell_length: edge_cell_length
        inv_dual_edge_length: inverse of dual_edge_length
        horizontal_start: horizontal start index
        horizontal_end: horizontal end index
    """

model/common/src/icon4py/model/common/metrics/metric_fields.py:607

• The docstring is incomplete. It should document the new parameters min_reduction and array_ns, as well as all existing parameters (flat_idx_max, e_owner_mask, lateral_boundary_level, nlev) and the return value.

    """
    compute the nflat_gradp value as the minimum value of the flat_idx_max array.
    """

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[jcanton]

LGTM

[github-actions]

Mandatory Tests

Please make sure you run these tests via comment before you merge!

• cscs-ci run default

Optional Tests

To run benchmarks you can use:

• cscs-ci run benchmark-bencher

To run tests and benchmarks with the DaCe backend you can use:

• cscs-ci run dace

To run test levels ignored by the default test suite (mostly simple datatest for static fields computations) you can use:

• cscs-ci run extra

For more detailed information please look at CI in the EXCLAIM universe.

[nfarabullini]

cscs-ci run default