pylibcudf.Column: add device_buffer_size and register a dask.sizeof function for cudf-polars Column and DataFrame

python/cudf_polars/cudf_polars/experimental/dask_serialize.py → python/cudf_polars/cudf_polars/experimental/dask_registers.py

@@ -1,12 +1,13 @@
 # SPDX-FileCopyrightText: Copyright (c) 2024-2025, NVIDIA CORPORATION & AFFILIATES.
 # SPDX-License-Identifier: Apache-2.0
 
-"""Dask serialization."""
+"""Dask function registrations such as serializers and dispatch implementations."""
 
 from __future__ import annotations
 
 from typing import TYPE_CHECKING, ClassVar, overload
 
+from dask.sizeof import sizeof as sizeof_dispatch
 from distributed.protocol import dask_deserialize, dask_serialize
 from distributed.protocol.cuda import cuda_deserialize, cuda_serialize
 from distributed.utils import log_errors
@@ -21,35 +22,35 @@
 
     from cudf_polars.typing import ColumnHeader, DataFrameHeader
 
-__all__ = ["SerializerManager", "register"]
+__all__ = ["DaskRegisterManager", "register"]
 
 
-class SerializerManager:  # pragma: no cover; Only used with Distributed scheduler
+class DaskRegisterManager:  # pragma: no cover; Only used with Distributed scheduler
     """Manager to ensure ensure serializer is only registered once."""
 
-    _serializer_registered: bool = False
+    _registered: bool = False
     _client_run_executed: ClassVar[set[str]] = set()
 
     @classmethod
-    def register_serialize(cls) -> None:
+    def register_once(cls) -> None:
         """Register Dask/cudf-polars serializers in calling process."""
-        if not cls._serializer_registered:
-            from cudf_polars.experimental.dask_serialize import register
+        if not cls._registered:
+            from cudf_polars.experimental.dask_registers import register
 
             register()
-            cls._serializer_registered = True
+            cls._registered = True
 
     @classmethod
     def run_on_cluster(cls, client: Client) -> None:
-        """Run serializer registration on the workers and scheduler."""
+        """Run register on the workers and scheduler once."""
         if client.id not in cls._client_run_executed:
-            client.run(cls.register_serialize)
-            client.run_on_scheduler(cls.register_serialize)
+            client.run(cls.register_once)
+            client.run_on_scheduler(cls.register_once)
             cls._client_run_executed.add(client.id)
 
 
 def register() -> None:
-    """Register dask serialization routines for DataFrames."""
+    """Register dask serialization and dispatch functions."""
 
     @overload
     def serialize_column_or_frame(
@@ -128,3 +129,20 @@ def _(header: ColumnHeader, frames: tuple[memoryview, memoryview]) -> Column:
             # Copy the second frame (the gpudata in host memory) back to the gpu
             frames = frames[0], plc.gpumemoryview(rmm.DeviceBuffer.to_device(frames[1]))
             return Column.deserialize(header, frames)
+
+    @sizeof_dispatch.register(Column)
+    def _(x: Column) -> int:
+        """The total size of the device buffers used by the Column."""
+        ret = 0
+        if x.obj.data() is not None:
+            ret += x.obj.data().nbytes
+        if x.obj.null_mask() is not None:
+            ret += x.obj.null_mask().nbytes
+        if x.obj.children() is not None:
+            ret += sum(sizeof_dispatch(c) for c in x.obj.children())
+        return ret
+
+    @sizeof_dispatch.register(DataFrame)
+    def _(x: DataFrame) -> int:
+        """The total size of the device buffers used by the DataFrame."""
+        return sum(sizeof_dispatch(c) for c in x.columns)

python/cudf_polars/cudf_polars/experimental/parallel.py

@@ -144,11 +144,11 @@ def get_scheduler(config_options: ConfigOptions) -> Any:
     ):  # pragma: no cover; block depends on executor type and Distributed cluster
         from distributed import get_client
 
-        from cudf_polars.experimental.dask_serialize import SerializerManager
+        from cudf_polars.experimental.dask_registers import DaskRegisterManager
 
         client = get_client()
-        SerializerManager.register_serialize()
-        SerializerManager.run_on_cluster(client)
+        DaskRegisterManager.register_once()
+        DaskRegisterManager.run_on_cluster(client)
         return client.get
     elif scheduler == "synchronous":
         from cudf_polars.experimental.scheduler import synchronous_scheduler

python/cudf_polars/tests/experimental/test_dask_serialize.py

@@ -13,7 +13,7 @@
 import rmm
 
 from cudf_polars.containers import DataFrame
-from cudf_polars.experimental.dask_serialize import register
+from cudf_polars.experimental.dask_registers import register
 
 # Must register serializers before running tests
 register()

python/cudf_polars/tests/experimental/test_dask_sizeof.py

@@ -0,0 +1,32 @@
+# SPDX-FileCopyrightText: Copyright (c) 2024-2025, NVIDIA CORPORATION & AFFILIATES.
+# SPDX-License-Identifier: Apache-2.0
+
+from __future__ import annotations
+
+import pyarrow as pa
+import pytest
+from dask.sizeof import sizeof
+
+import pylibcudf as plc
+
+from cudf_polars.containers import DataFrame
+from cudf_polars.experimental.dask_registers import register
+
+# Must register sizeof dispatch before running tests
+register()
+
+
+@pytest.mark.parametrize(
+    "arrow_tbl, size",
+    [
+        (pa.table([]), 0),
+        (pa.table({"a": [1, 2, 3], "b": [4, 5, 6], "c": [7, 8, 9]}), 9 * 8),
+        (pa.table({"a": [1, 2, 3]}), 3 * 8),
+        (pa.table({"a": [1], "b": [2], "c": [3]}), 3 * 8),
+        (pa.table({"a": [1, 2, None]}), 88),
+    ],
+)
+def test_dask_sizeof(arrow_tbl, size):
+    plc_tbl = plc.interop.from_arrow(arrow_tbl)
+    df = DataFrame.from_table(plc_tbl, names=arrow_tbl.column_names)
+    assert sizeof(df) == size