Connectivity peakmem benchmarks

benchmarks/bench_connectivity.py

@@ -2,8 +2,12 @@
 import urllib.request
 from pathlib import Path
 
+import numpy as np
+
 import uxarray as ux
 
+from .helpers._peakmem import peak_allocated
+
 current_path = Path(os.path.dirname(os.path.realpath(__file__)))
 
 grid_filename_480 = "oQU480.grid.nc"
@@ -63,79 +67,297 @@ def teardown(self, resolution, *args, **kwargs):
     "node_face_connectivity",
 ]
 
+# Direct prerequisites only, read off the ``_populate_*`` functions in
+# ``uxarray/grid/connectivity.py``; accessing one builds its own in turn.
+CONNECTIVITY_PREREQUISITES = {
+    "n_nodes_per_face": (),
+    "face_node_connectivity": (),
+    "edge_node_connectivity": ("n_nodes_per_face",),
+    # ``_populate_edge_node_connectivity`` writes this one out too, so it costs
+    # nothing once that has run.
+    "face_edge_connectivity": ("edge_node_connectivity",),
+    "node_edge_connectivity": ("edge_node_connectivity",),
+    "face_face_connectivity": ("edge_face_connectivity",),
+    "edge_face_connectivity": ("face_edge_connectivity",),
+    "node_face_connectivity": (),
+}
+
+
+def _build_prerequisites(uxgrid, connectivity):
+    """Builds ``connectivity``'s prerequisites, so what follows measures one
+    construction routine rather than the whole chain rooted at it."""
+    for prerequisite in CONNECTIVITY_PREREQUISITES[connectivity]:
+        getattr(uxgrid, prerequisite)
+    return uxgrid
+
+
 _numba_warmed_up = False
 
-def _warmup(uxgrid):
+def _warmup():
     """Compiles the Numba kernels backing each connectivity variable.
 
-    ``_build_node_edge_connectivity`` is not disk-cached, so a fresh benchmark
-    process would otherwise charge ~240ms of JIT compilation to whichever sample
-    happened to touch it first.
+    Every kernel in ``uxarray/grid/connectivity.py`` is ``@njit(cache=True)``, so
+    this carries across processes through Numba's on-disk cache -- what makes it
+    usable from ``setup_cache``. Loading from that cache still allocates, so it
+    matters for ``track_peakmem_*`` too, not just timing.
     """
     global _numba_warmed_up
     if _numba_warmed_up:
         return
+    # Resolution affects how long the kernels run, not which signatures compile.
+    uxgrid = ux.Grid.from_topology(*_source_topology(GridBenchmark.params[0][0]))
     for name in CONNECTIVITY_NAMES:
         getattr(uxgrid, name)
     _numba_warmed_up = True
 
 
-class Connectivity(GridBenchmark):
-    # Each connectivity variable is cached in ``Grid._ds`` once constructed, so a
-    # sample may only contain a single call; otherwise every call but the first
-    # would time a dictionary lookup.
-    number = 1
+_topology_cache = {}
 
-    def setup(self, resolution, *args, **kwargs):
-        # The benchmark grids are MPAS meshes, which carry every connectivity
-        # variable on disk. Reading one would time the MPAS parser rather than
-        # the construction routines, so reduce the grid down to the minimal
-        # UGRID topology and let each variable be built on demand.
+
+def _source_topology(resolution):
+    """The minimal UGRID topology for ``resolution``, read once per process.
+
+    The benchmark grids are MPAS meshes carrying every connectivity variable on
+    disk; reading one would measure the MPAS parser rather than the construction
+    routines, so each variable is left to be built on demand.
+
+    Cached because asv re-runs ``setup`` between repeats, and at dyamond
+    resolutions re-reading the source grid dwarfs the sample it precedes.
+    """
+    if resolution not in _topology_cache:
         source_grid = ux.open_grid(file_path_dict[resolution])
-        self.topology = (
+        _topology_cache[resolution] = (
             source_grid.node_lon.data,
             source_grid.node_lat.data,
             source_grid.face_node_connectivity.data,
         )
+    return _topology_cache[resolution]
+
 
-        _warmup(self.minimal_grid())
+class MinimalGridBenchmark(GridBenchmark):
+    """Template for benchmarks that construct connectivity variables on demand.
+
+    Holds a ``Grid`` carrying nothing but the minimal UGRID topology, plus the
+    topology needed to mint further ones, and leaves the Numba kernels compiled.
+    """
+
+    # Handover slot for ``_prerequisite_setup``; see its docstring.
+    active_grid = None
+
+    # asv's 60s default is not enough to build a connectivity variable at 3.75km.
+    timeout = 1200
+
+    def setup(self, resolution, *args, **kwargs):
+        self.topology = _source_topology(resolution)
+
+        _warmup()
         self.uxgrid = self.minimal_grid()
+        MinimalGridBenchmark.active_grid = self.uxgrid
 
     def minimal_grid(self):
+        """Mints a ``Grid`` holding nothing beyond the minimal UGRID topology."""
         return ux.Grid.from_topology(*self.topology)
 
     def teardown(self, resolution, *args, **kwargs):
+        # Cleared so a per-benchmark setup raises rather than quietly measuring
+        # a stale grid.
+        MinimalGridBenchmark.active_grid = None
         del self.uxgrid
         del self.topology
 
+
+def _prerequisite_setup(connectivity):
+    """Builds a per-benchmark ``setup`` that puts ``connectivity``'s
+    prerequisites in place before the clock starts.
+
+    asv collects ``setup`` from the benchmark function as well as the class and
+    runs the class one first, but passes neither the instance, hence the handover
+    through ``MinimalGridBenchmark.active_grid``.
+    """
+
+    def setup(resolution, *args, **kwargs):
+        _build_prerequisites(MinimalGridBenchmark.active_grid, connectivity)
+
+    return setup
+
+
+class Connectivity(MinimalGridBenchmark):
+    """Time to construct each connectivity variable.
+
+    Prerequisites are built during ``setup``, so a sample times the one routine
+    that produces that variable rather than the whole chain rooted at it --
+    matching how :class:`ConnectivityTracemalloc` attributes memory.
+    """
+
+    number = 1
+    warmup_time = 0
+
     def time_n_nodes_per_face(self, resolution):
         _ = self.uxgrid.n_nodes_per_face.compute()
 
+    time_n_nodes_per_face.setup = _prerequisite_setup("n_nodes_per_face")
+
     def time_face_node(self, resolution):
         _ = self.uxgrid.face_node_connectivity.compute()
 
+    time_face_node.setup = _prerequisite_setup("face_node_connectivity")
+
     def time_edge_node(self, resolution):
         _ = self.uxgrid.edge_node_connectivity.compute()
 
+    time_edge_node.setup = _prerequisite_setup("edge_node_connectivity")
+
 #   TODO: Not yet supported?
 #   def time_node_node(self, resolution):
 #       _ = self.uxgrid.node_node_connectivity
 
     def time_face_edge(self, resolution):
         _ = self.uxgrid.face_edge_connectivity.compute()
 
+    time_face_edge.setup = _prerequisite_setup("face_edge_connectivity")
+
 #   TODO: Not yet supported?
 #   def time_edge_edge(self, resolution):
 #        _ = self.uxgrid.edge_edge_connectivity
 
     def time_node_edge(self, resolution):
         _ = self.uxgrid.node_edge_connectivity.compute()
 
+    time_node_edge.setup = _prerequisite_setup("node_edge_connectivity")
+
     def time_face_face(self, resolution):
         _ = self.uxgrid.face_face_connectivity.compute()
 
+    time_face_face.setup = _prerequisite_setup("face_face_connectivity")
+
     def time_edge_face(self, resolution):
         _ = self.uxgrid.edge_face_connectivity.compute()
 
+    time_edge_face.setup = _prerequisite_setup("edge_face_connectivity")
+
     def time_node_face(self, resolution):
         _ = self.uxgrid.node_face_connectivity.compute()
+
+    time_node_face.setup = _prerequisite_setup("node_face_connectivity")
+
+
+class ConnectivityTracemalloc(MinimalGridBenchmark):
+    """Peak memory of each connectivity routine on its own.
+
+    The transient high-water allocation of the construction routine, with the
+    ~245MB the process already holds excluded.
+    """
+
+    unit = "bytes"
+
+    def _peak_building(self, name):
+        """Peak allocation of ``name``'s own construction routine."""
+        uxgrid = _build_prerequisites(self.minimal_grid(), name)
+        return peak_allocated(lambda: getattr(uxgrid, name).compute())
+
+    def track_peakmem_n_nodes_per_face(self, resolution):
+        return self._peak_building("n_nodes_per_face")
+
+    def track_peakmem_face_node(self, resolution):
+        return self._peak_building("face_node_connectivity")
+
+    def track_peakmem_edge_node(self, resolution):
+        return self._peak_building("edge_node_connectivity")
+
+    def track_peakmem_face_edge(self, resolution):
+        return self._peak_building("face_edge_connectivity")
+
+    def track_peakmem_node_edge(self, resolution):
+        return self._peak_building("node_edge_connectivity")
+
+    def track_peakmem_face_face(self, resolution):
+        return self._peak_building("face_face_connectivity")
+
+    def track_peakmem_edge_face(self, resolution):
+        return self._peak_building("edge_face_connectivity")
+
+    def track_peakmem_node_face(self, resolution):
+        return self._peak_building("node_face_connectivity")
+
+
+def _save_topology(uxgrid, npz_path):
+    """Writes the minimal UGRID topology of ``uxgrid`` out to ``npz_path``.
+
+    Compressed because ``setup_cache`` writes into a ``tempfile.mkdtemp()`` and
+    ``face_node_connectivity`` alone is multi-GB at 3.75km.
+    """
+    np.savez_compressed(
+        npz_path,
+        node_lon=uxgrid.node_lon.data,
+        node_lat=uxgrid.node_lat.data,
+        face_node_connectivity=uxgrid.face_node_connectivity.data,
+    )
+
+
+def _load_topology(npz_path):
+    """Builds a ``Grid`` holding nothing beyond the minimal UGRID topology."""
+    with np.load(npz_path) as topology:
+        return ux.Grid.from_topology(
+            topology["node_lon"],
+            topology["node_lat"],
+            topology["face_node_connectivity"],
+        )
+
+
+class ConnectivityChainRss:
+    """Peak resident memory of the process while constructing each connectivity
+    variable.
+
+    Differs from :class:`ConnectivityTracemalloc` on two axes. Scope: no
+    prerequisites are built in ``setup``, so a sample covers
+    the whole chain rooted at that variable. Instrument: ``ru_maxrss`` for the
+    whole process, so the ~250MB import.
+    """
+
+    # Only the parameterization is shared with ``MinimalGridBenchmark``
+    param_names = GridBenchmark.param_names
+    params = GridBenchmark.params
+    timeout = 1200
+
+    def setup_cache(self):
+        topology_paths = {}
+        for resolution in self.params[0]:
+            npz_path = os.path.abspath(f"topology_{resolution}.npz")
+            _save_topology(ux.open_grid(file_path_dict[resolution]), npz_path)
+            topology_paths[resolution] = npz_path
+
+        _warmup()
+
+        return topology_paths
+
+    setup_cache.timeout = 1800
+
+    def setup(self, topology_paths, resolution):
+        self.uxgrid = _load_topology(topology_paths[resolution])
+
+    def teardown(self, topology_paths, resolution):
+        del self.uxgrid
+
+    def peakmem_n_nodes_per_face(self, topology_paths, resolution):
+        _ = self.uxgrid.n_nodes_per_face.compute()
+
+    def peakmem_face_node(self, topology_paths, resolution):
+        _ = self.uxgrid.face_node_connectivity.compute()
+
+    def peakmem_edge_node(self, topology_paths, resolution):
+        _ = self.uxgrid.edge_node_connectivity.compute()
+
+    def peakmem_face_edge(self, topology_paths, resolution):
+        _ = self.uxgrid.face_edge_connectivity.compute()
+
+    def peakmem_node_edge(self, topology_paths, resolution):
+        _ = self.uxgrid.node_edge_connectivity.compute()
+
+    def peakmem_face_face(self, topology_paths, resolution):
+        _ = self.uxgrid.face_face_connectivity.compute()
+
+    def peakmem_edge_face(self, topology_paths, resolution):
+        _ = self.uxgrid.edge_face_connectivity.compute()
+
+    def peakmem_node_face(self, topology_paths, resolution):
+        _ = self.uxgrid.node_face_connectivity.compute()

uxarray/grid/connectivity.py

@@ -469,7 +469,7 @@ def _populate_node_edge_connectivity(grid):
     )
 
 
-@njit
+@njit(cache=True)
 def _build_node_edge_connectivity(edge_nodes, n_node):
     """Constructs the Node Edge Connectivity, which stores the indices of the edges that are shared by each node."""
     n_edge, nodes_per_edge = edge_nodes.shape