[Superseded] Spilling to host memory seamlessly

ci/gpu/build.sh

@@ -197,7 +197,7 @@ else
     # copied by CI from the upstream 11.5 jobs into $CONDA_ARTIFACT_PATH
     gpuci_logger "Installing cudf, dask-cudf, cudf_kafka, and custreamz"
     gpuci_mamba_retry install cudf dask-cudf cudf_kafka custreamz -c "${CONDA_BLD_DIR}" -c "${CONDA_ARTIFACT_PATH}"
-    
+
     gpuci_logger "Check current conda environment"
     conda list --show-channel-urls
 
@@ -282,6 +282,10 @@ conda list
 gpuci_logger "Python py.test for cuDF"
 py.test -n 8 --cache-clear --basetemp="$WORKSPACE/cudf-cuda-tmp" --ignore="$WORKSPACE/python/cudf/cudf/benchmarks" --junitxml="$WORKSPACE/junit-cudf.xml" -v --cov-config="$WORKSPACE/python/cudf/.coveragerc" --cov=cudf --cov-report=xml:"$WORKSPACE/python/cudf/cudf-coverage.xml" --cov-report term --dist=loadscope tests
 
+gpuci_logger "Python py.tests for cuDF with spilling (CUDF_SPILL_DEVICE_LIMIT=1)"
+# Due to time concerns, we only run a limited set of tests
+CUDF_SPILL=on CUDF_SPILL_DEVICE_LIMIT=1 py.test -n 8 --cache-clear --basetemp="$WORKSPACE/cudf-cuda-tmp" --ignore="$WORKSPACE/python/cudf/cudf/benchmarks" -v --cov-config="$WORKSPACE/python/cudf/.coveragerc" --cov-append --cov=cudf --cov-report=xml:"$WORKSPACE/python/cudf/cudf-coverage.xml" --cov-report term --dist=loadscope tests/test_binops.py tests/test_dataframe.py tests/test_buffer.py tests/test_onehot.py tests/test_reshape.py
+
 cd "$WORKSPACE/python/dask_cudf"
 gpuci_logger "Python py.test for dask-cudf"
 py.test -n 8 --cache-clear --basetemp="$WORKSPACE/dask-cudf-cuda-tmp" --junitxml="$WORKSPACE/junit-dask-cudf.xml" -v --cov-config=.coveragerc --cov=dask_cudf --cov-report=xml:"$WORKSPACE/python/dask_cudf/dask-cudf-coverage.xml" --cov-report term dask_cudf
@@ -290,7 +294,6 @@ cd "$WORKSPACE/python/custreamz"
 gpuci_logger "Python py.test for cuStreamz"
 py.test -n 8 --cache-clear --basetemp="$WORKSPACE/custreamz-cuda-tmp" --junitxml="$WORKSPACE/junit-custreamz.xml" -v --cov-config=.coveragerc --cov=custreamz --cov-report=xml:"$WORKSPACE/python/custreamz/custreamz-coverage.xml" --cov-report term custreamz
 
-
 # only install strings_udf after cuDF is finished testing without its presence
 gpuci_logger "Installing strings_udf"
 gpuci_mamba_retry install strings_udf -c "${CONDA_BLD_DIR}" -c "${CONDA_ARTIFACT_PATH}"

docs/cudf/source/developer_guide/library_design.md

@@ -27,24 +27,24 @@ Finally we tie these pieces together to provide a more holistic view of the proj
 % class RangeIndex
 % class DataFrame
 % class Series
-% 
+%
 % Frame <|-- IndexedFrame
-% 
+%
 % Frame <|-- SingleColumnFrame
-% 
+%
 % SingleColumnFrame <|-- Series
 % IndexedFrame <|-- Series
-% 
+%
 % IndexedFrame <|-- DataFrame
-% 
+%
 % BaseIndex <|-- RangeIndex
-% 
+%
 % BaseIndex <|-- MultiIndex
 % Frame <|-- MultiIndex
-% 
+%
 % BaseIndex <|-- GenericIndex
 % SingleColumnFrame <|-- GenericIndex
-% 
+%
 % @enduml
 
 
@@ -212,6 +212,12 @@ Conversely, when constructed from a host object,
 The data is then copied from the host object into the newly allocated device memory.
 You can read more about [device memory allocation with RMM here](https://github.com/rapidsai/rmm).
 
+### Spilling to host memory
+
+Setting the environment variable `CUDF_SPILL=on` enables automatic spilling (and "unspilling") of buffers from
+device to host to enable out-of-memory computation, i.e., computing on objects that occupy more memory than is
+available on the GPU.
+
 ## The Cython layer
 
 The lowest level of cuDF is its interaction with `libcudf` via Cython.

python/cudf/cudf/_lib/binaryop.pyx

@@ -25,9 +25,11 @@ from cudf._lib.cpp.types cimport data_type, type_id
 from cudf._lib.types cimport dtype_to_data_type, underlying_type_t_type_id
 
 from cudf.api.types import is_scalar, is_string_dtype
+from cudf.core.buffer.spillable_buffer import SpillLock
 
 cimport cudf._lib.cpp.binaryop as cpp_binaryop
 from cudf._lib.cpp.binaryop cimport binary_operator
+
 import cudf
 
 
@@ -102,9 +104,9 @@ class BinaryOperation(IntEnum):
 
 cdef binaryop_v_v(Column lhs, Column rhs,
                   binary_operator c_op, data_type c_dtype):
-    cdef column_view c_lhs = lhs.view()
-    cdef column_view c_rhs = rhs.view()
-
+    slock = SpillLock()
+    cdef column_view c_lhs = lhs.view(slock)
+    cdef column_view c_rhs = rhs.view(slock)
     cdef unique_ptr[column] c_result
 
     with nogil:
@@ -122,7 +124,8 @@ cdef binaryop_v_v(Column lhs, Column rhs,
 
 cdef binaryop_v_s(Column lhs, DeviceScalar rhs,
                   binary_operator c_op, data_type c_dtype):
-    cdef column_view c_lhs = lhs.view()
+    slock = SpillLock()
+    cdef column_view c_lhs = lhs.view(slock)
     cdef const scalar* c_rhs = rhs.get_raw_ptr()
 
     cdef unique_ptr[column] c_result
@@ -142,7 +145,8 @@ cdef binaryop_v_s(Column lhs, DeviceScalar rhs,
 cdef binaryop_s_v(DeviceScalar lhs, Column rhs,
                   binary_operator c_op, data_type c_dtype):
     cdef const scalar* c_lhs = lhs.get_raw_ptr()
-    cdef column_view c_rhs = rhs.view()
+    slock = SpillLock()
+    cdef column_view c_rhs = rhs.view(slock)
 
     cdef unique_ptr[column] c_result
 
@@ -213,8 +217,9 @@ def binaryop_udf(Column lhs, Column rhs, udf_ptx, dtype):
     has to be specified in `dtype`, a numpy data type.
     Currently ONLY int32, int64, float32 and float64 are supported.
     """
-    cdef column_view c_lhs = lhs.view()
-    cdef column_view c_rhs = rhs.view()
+    slock = SpillLock()
+    cdef column_view c_lhs = lhs.view(slock)
+    cdef column_view c_rhs = rhs.view(slock)
 
     cdef type_id tid = (
         <type_id> (

python/cudf/cudf/_lib/column.pxd

@@ -1,4 +1,4 @@
-# Copyright (c) 2020, NVIDIA CORPORATION.
+# Copyright (c) 2020-2022, NVIDIA CORPORATION.
 
 from libcpp cimport bool
 from libcpp.memory cimport unique_ptr
@@ -23,12 +23,16 @@ cdef class Column:
         cdef object _mask
         cdef object _null_count
 
-    cdef column_view _view(self, size_type null_count) except *
-    cdef column_view view(self) except *
+    cdef column_view _view(
+        self, size_type null_count, spill_lock
+    ) except *
+    cdef column_view view(self, spill_lock=*) except *
     cdef mutable_column_view mutable_view(self) except *
 
     @staticmethod
-    cdef Column from_unique_ptr(unique_ptr[column] c_col)
+    cdef Column from_unique_ptr(
+        unique_ptr[column] c_col, bint data_ptr_exposed=*
+    )
 
     @staticmethod
     cdef Column from_column_view(column_view, object)