Sperf & Cperf Benchmarks (#4621)

benchmarks/README.md

@@ -21,13 +21,20 @@ automated overnight run locally. See the session docstring for detail.
 
 ### Environment variables
 
-* ``DATA_GEN_PYTHON`` - required - path to a Python executable that can be
+* `OVERRIDE_TEST_DATA_REPOSITORY` - required - some benchmarks use
+`iris-test-data` content, and your local `site.cfg` is not available for
+benchmark scripts.
+* `DATA_GEN_PYTHON` - required - path to a Python executable that can be
 used to generate benchmark test objects/files; see
 [Data generation](#data-generation). The Nox session sets this automatically,
 but will defer to any value already set in the shell.
-* ``BENCHMARK_DATA`` - optional - path to a directory for benchmark synthetic
+* `BENCHMARK_DATA` - optional - path to a directory for benchmark synthetic
 test data, which the benchmark scripts will create if it doesn't already
-exist. Defaults to ``<root>/benchmarks/.data/`` if not set.
+exist. Defaults to `<root>/benchmarks/.data/` if not set.
+* `ON_DEMAND_BENCHMARKS` - optional - when set (to any value): benchmarks
+decorated with `@on_demand_benchmark` are included in the ASV run. Usually
+coupled with the ASV `--bench` argument to only run the benchmark(s) of
+interest. Is set during the Nox `cperf` and `sperf` sessions.
 
 ## Writing benchmarks
 
@@ -65,6 +72,16 @@ be significantly larger (e.g. a 1000x1000 `Cube`). Performance differences
 might only be seen for the larger value, or the smaller, or both, getting you
 closer to the root cause.
 
+### On-demand benchmarks
+
+Some benchmarks provide useful insight but are inappropriate to be included in
+a benchmark run by default, e.g. those with long run-times or requiring a local
+file. These benchmarks should be decorated with `@on_demand_benchmark`
+(see [benchmarks init](./benchmarks/__init__.py)), which
+sets the benchmark to only be included in a run when the `ON_DEMAND_BENCHMARKS`
+environment variable is set. Examples include the CPerf and SPerf benchmark
+suites for the UK Met Office NG-VAT project.
+
 ## Benchmark environments
 
 We have disabled ASV's standard environment management, instead using an

benchmarks/asv.conf.json

@@ -5,6 +5,7 @@
     "repo": "..",
     "environment_type": "conda-delegated",
     "show_commit_url": "http://github.com/scitools/iris/commit/",
+    "branches": ["upstream/main"],
 
     "benchmark_dir": "./benchmarks",
     "env_dir": ".asv/env",

benchmarks/benchmarks/__init__.py

@@ -4,10 +4,10 @@
 # See COPYING and COPYING.LESSER in the root of the repository for full
 # licensing details.
 """Common code for benchmarks."""
+from functools import wraps
+from os import environ
 import resource
 
-from .generate_data import BENCHMARK_DATA, run_function_elsewhere
-
 ARTIFICIAL_DIM_SIZE = int(10e3)  # For all artificial cubes, coords etc.
 
 
@@ -70,3 +70,60 @@ def __exit__(self, *_):
     def addedmem_mb(self):
         """Return measured memory growth, in Mb."""
         return self.mb_after - self.mb_before
+
+    @staticmethod
+    def decorator(changed_params: list = None):
+        """
+        Decorates this benchmark to track growth in resident memory during execution.
+
+        Intended for use on ASV ``track_`` benchmarks. Applies the
+        :class:`TrackAddedMemoryAllocation` context manager to the benchmark
+        code, sets the benchmark ``unit`` attribute to ``Mb``. Optionally
+        replaces the benchmark ``params`` attribute with ``changed_params`` -
+        useful to avoid testing very small memory volumes, where the results
+        are vulnerable to noise.
+
+        Parameters
+        ----------
+        changed_params : list
+            Replace the benchmark's ``params`` attribute with this list.
+
+        """
+        if changed_params:
+            # Must make a copy for re-use safety!
+            _changed_params = list(changed_params)
+        else:
+            _changed_params = None
+
+        def _inner_decorator(decorated_func):
+            @wraps(decorated_func)
+            def _inner_func(*args, **kwargs):
+                assert decorated_func.__name__[:6] == "track_"
+                # Run the decorated benchmark within the added memory context manager.
+                with TrackAddedMemoryAllocation() as mb:
+                    decorated_func(*args, **kwargs)
+                return mb.addedmem_mb()
+
+            if _changed_params:
+                # Replace the params if replacement provided.
+                _inner_func.params = _changed_params
+            _inner_func.unit = "Mb"
+            return _inner_func
+
+        return _inner_decorator
+
+
+def on_demand_benchmark(benchmark_object):
+    """
+    Decorator. Disables these benchmark(s) unless ON_DEMAND_BENCHARKS env var is set.
+
+    For benchmarks that, for whatever reason, should not be run by default.
+    E.g:
+        * Require a local file
+        * Used for scalability analysis instead of commit monitoring.
+
+    Can be applied to benchmark classes/methods/functions.
+
+    """
+    if "ON_DEMAND_BENCHMARKS" in environ:
+        return benchmark_object

benchmarks/benchmarks/cperf/__init__.py

@@ -0,0 +1,97 @@
+# Copyright Iris contributors
+#
+# This file is part of Iris and is released under the LGPL license.
+# See COPYING and COPYING.LESSER in the root of the repository for full
+# licensing details.
+"""
+Benchmarks for the CPerf scheme of the UK Met Office's NG-VAT project.
+
+CPerf = comparing performance working with data in UM versus LFRic formats.
+
+Files available from the UK Met Office:
+  moo ls moose:/adhoc/projects/avd/asv/data_for_nightly_tests/
+"""
+import numpy as np
+
+from iris import load_cube
+
+# TODO: remove uses of PARSE_UGRID_ON_LOAD once UGRID parsing is core behaviour.
+from iris.experimental.ugrid import PARSE_UGRID_ON_LOAD
+
+from ..generate_data import BENCHMARK_DATA
+from ..generate_data.ugrid import make_cubesphere_testfile
+
+# The data of the core test UM files has dtype=np.float32 shape=(1920, 2560)
+_UM_DIMS_YX = (1920, 2560)
+# The closest cubesphere size in terms of datapoints is sqrt(1920*2560 / 6)
+#  This gives ~= 905, i.e. "C905"
+_N_CUBESPHERE_UM_EQUIVALENT = int(np.sqrt(np.prod(_UM_DIMS_YX) / 6))
+
+
+class SingleDiagnosticMixin:
+    """For use in any benchmark classes that work on a single diagnostic file."""
+
+    params = [
+        ["LFRic", "UM", "UM_lbpack0", "UM_netcdf"],
+        [False, True],
+        [False, True],
+    ]
+    param_names = ["file type", "height dim (len 71)", "time dim (len 3)"]
+
+    def setup(self, file_type, three_d, three_times):
+        if file_type == "LFRic":
+            # Generate an appropriate synthetic LFRic file.
+            if three_times:
+                n_times = 3
+            else:
+                n_times = 1
+
+            # Use a cubesphere size ~equivalent to our UM test data.
+            cells_per_panel_edge = _N_CUBESPHERE_UM_EQUIVALENT
+            create_kwargs = dict(c_size=cells_per_panel_edge, n_times=n_times)
+
+            if three_d:
+                create_kwargs["n_levels"] = 71
+
+            # Will re-use a file if already present.
+            file_path = make_cubesphere_testfile(**create_kwargs)
+
+        else:
+            # Locate the appropriate UM file.
+            if three_times:
+                # pa/pb003 files
+                numeric = "003"
+            else:
+                # pa/pb000 files
+                numeric = "000"
+
+            if three_d:
+                # theta diagnostic, N1280 file w/ 71 levels (1920, 2560, 71)
+                file_name = f"umglaa_pb{numeric}-theta"
+            else:
+                # surface_temp diagnostic, N1280 file (1920, 2560)
+                file_name = f"umglaa_pa{numeric}-surfacetemp"
+
+            file_suffices = {
+                "UM": "",  # packed FF (WGDOS lbpack = 1)
+                "UM_lbpack0": ".uncompressed",  # unpacked FF (lbpack = 0)
+                "UM_netcdf": ".nc",  # UM file -> Iris -> NetCDF file
+            }
+            suffix = file_suffices[file_type]
+
+            file_path = (BENCHMARK_DATA / file_name).with_suffix(suffix)
+            if not file_path.exists():
+                message = "\n".join(
+                    [
+                        f"Expected local file not found: {file_path}",
+                        "Available from the UK Met Office.",
+                    ]
+                )
+                raise FileNotFoundError(message)
+
+        self.file_path = file_path
+        self.file_type = file_type
+
+    def load(self):
+        with PARSE_UGRID_ON_LOAD.context():
+            return load_cube(str(self.file_path))

benchmarks/benchmarks/cperf/equality.py

@@ -0,0 +1,58 @@
+# Copyright Iris contributors
+#
+# This file is part of Iris and is released under the LGPL license.
+# See COPYING and COPYING.LESSER in the root of the repository for full
+# licensing details.
+"""
+Equality benchmarks for the CPerf scheme of the UK Met Office's NG-VAT project.
+"""
+from . import SingleDiagnosticMixin
+from .. import on_demand_benchmark
+
+
+class EqualityMixin(SingleDiagnosticMixin):
+    """
+    Uses :class:`SingleDiagnosticMixin` as the realistic case will be comparing
+    :class:`~iris.cube.Cube`\\ s that have been loaded from file.
+    """
+
+    # Cut down the parent parameters.
+    params = [["LFRic", "UM"]]
+
+    def setup(self, file_type, three_d=False, three_times=False):
+        super().setup(file_type, three_d, three_times)
+        self.cube = self.load()
+        self.other_cube = self.load()
+
+
+@on_demand_benchmark
+class CubeEquality(EqualityMixin):
+    """
+    Benchmark time and memory costs of comparing LFRic and UM
+     :class:`~iris.cube.Cube`\\ s.
+    """
+
+    def _comparison(self):
+        _ = self.cube == self.other_cube
+
+    def peakmem_eq(self, file_type):
+        self._comparison()
+
+    def time_eq(self, file_type):
+        self._comparison()
+
+
+@on_demand_benchmark
+class MeshEquality(EqualityMixin):
+    """Provides extra context for :class:`CubeEquality`."""
+
+    params = [["LFRic"]]
+
+    def _comparison(self):
+        _ = self.cube.mesh == self.other_cube.mesh
+
+    def peakmem_eq(self, file_type):
+        self._comparison()
+
+    def time_eq(self, file_type):
+        self._comparison()

benchmarks/benchmarks/cperf/load.py

@@ -0,0 +1,57 @@
+# Copyright Iris contributors
+#
+# This file is part of Iris and is released under the LGPL license.
+# See COPYING and COPYING.LESSER in the root of the repository for full
+# licensing details.
+"""
+File loading benchmarks for the CPerf scheme of the UK Met Office's NG-VAT project.
+"""
+from . import SingleDiagnosticMixin
+from .. import on_demand_benchmark
+
+
+@on_demand_benchmark
+class SingleDiagnosticLoad(SingleDiagnosticMixin):
+    def time_load(self, _, __, ___):
+        """
+        The 'real world comparison'
+          * UM coords are always realised (DimCoords).
+          * LFRic coords are not realised by default (MeshCoords).
+
+        """
+        cube = self.load()
+        assert cube.has_lazy_data()
+        # UM files load lon/lat as DimCoords, which are always realised.
+        expecting_lazy_coords = self.file_type == "LFRic"
+        for coord_name in "longitude", "latitude":
+            coord = cube.coord(coord_name)
+            assert coord.has_lazy_points() == expecting_lazy_coords
+            assert coord.has_lazy_bounds() == expecting_lazy_coords
+
+    def time_load_w_realised_coords(self, _, __, ___):
+        """A valuable extra comparison where both UM and LFRic coords are realised."""
+        cube = self.load()
+        for coord_name in "longitude", "latitude":
+            coord = cube.coord(coord_name)
+            # Don't touch actual points/bounds objects - permanent
+            #  realisation plays badly with ASV's re-run strategy.
+            if coord.has_lazy_points():
+                coord.core_points().compute()
+            if coord.has_lazy_bounds():
+                coord.core_bounds().compute()
+
+
+@on_demand_benchmark
+class SingleDiagnosticRealise(SingleDiagnosticMixin):
+    # The larger files take a long time to realise.
+    timeout = 600.0
+
+    def setup(self, file_type, three_d, three_times):
+        super().setup(file_type, three_d, three_times)
+        self.loaded_cube = self.load()
+
+    def time_realise(self, _, __, ___):
+        # Don't touch loaded_cube.data - permanent realisation plays badly with
+        #  ASV's re-run strategy.
+        assert self.loaded_cube.has_lazy_data()
+        self.loaded_cube.core_data().compute()

benchmarks/benchmarks/cperf/save.py

@@ -0,0 +1,47 @@
+# Copyright Iris contributors
+#
+# This file is part of Iris and is released under the LGPL license.
+# See COPYING and COPYING.LESSER in the root of the repository for full
+# licensing details.
+"""
+File saving benchmarks for the CPerf scheme of the UK Met Office's NG-VAT project.
+"""
+
+from iris import save
+
+from . import _N_CUBESPHERE_UM_EQUIVALENT, _UM_DIMS_YX
+from .. import TrackAddedMemoryAllocation, on_demand_benchmark
+from ..generate_data.ugrid import (
+    make_cube_like_2d_cubesphere,
+    make_cube_like_umfield,
+)
+
+
+@on_demand_benchmark
+class NetcdfSave:
+    """
+    Benchmark time and memory costs of saving ~large-ish data cubes to netcdf.
+    Parametrised by file type.
+
+    """
+
+    params = ["LFRic", "UM"]
+    param_names = ["data type"]
+
+    def setup(self, data_type):
+        if data_type == "LFRic":
+            self.cube = make_cube_like_2d_cubesphere(
+                n_cube=_N_CUBESPHERE_UM_EQUIVALENT, with_mesh=True
+            )
+        else:
+            self.cube = make_cube_like_umfield(_UM_DIMS_YX)
+
+    def _save_data(self, cube):
+        save(cube, "tmp.nc")
+
+    def time_save_data_netcdf(self, data_type):
+        self._save_data(self.cube)
+
+    @TrackAddedMemoryAllocation.decorator()
+    def track_addedmem_save_data_netcdf(self, data_type):
+        self._save_data(self.cube)

benchmarks/benchmarks/experimental/ugrid.py → benchmarks/benchmarks/experimental/ugrid/__init__.py

@@ -14,8 +14,8 @@
 
 from iris.experimental import ugrid
 
-from .. import ARTIFICIAL_DIM_SIZE, disable_repeat_between_setup
-from ..generate_data.stock import sample_mesh
+from ... import ARTIFICIAL_DIM_SIZE, disable_repeat_between_setup
+from ...generate_data.stock import sample_mesh
 
 
 class UGridCommon: