Fallback resource adaptor

include/rmm/mr/device/failure_alternate_resource_adaptor.hpp

@@ -0,0 +1,159 @@
+/*
+ * Copyright (c) 2024, NVIDIA CORPORATION.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#pragma once
+
+#include <rmm/detail/error.hpp>
+#include <rmm/detail/export.hpp>
+#include <rmm/mr/device/device_memory_resource.hpp>
+#include <rmm/resource_ref.hpp>
+
+#include <cstddef>
+#include <mutex>
+#include <unordered_set>
+
+namespace RMM_NAMESPACE {
+namespace mr {
+/**
+ * @addtogroup device_resource_adaptors
+ * @{
+ * @file
+ */
+
+/**
+ * @brief A device memory resource that uses an alternate upstream resource when the primary
+ * upstream resource throws a specified exception type.
+ *
+ * An instance of this resource must be constructed with two upstream resources to satisfy
+ * allocation requests.
+ *
+ * @tparam ExceptionType The type of exception that this adaptor should respond to.
+ */
+template <typename ExceptionType = rmm::out_of_memory>
+class failure_alternate_resource_adaptor final : public device_memory_resource {
+ public:
+  using exception_type = ExceptionType;  ///< The type of exception this object catches/throws
+
+  /**
+   * @brief Construct a new `failure_alternate_resource_adaptor` that uses `primary_upstream`
+   * to satisfy allocation requests and if that fails with `ExceptionType`, uses
+   * `alternate_upstream`.
+   *
+   * @param primary_upstream The primary resource used for allocating/deallocating device memory
+   * @param alternate_upstream The alternate resource used for allocating/deallocating device memory
+   * memory
+   */
+  failure_alternate_resource_adaptor(device_async_resource_ref primary_upstream,
+                                     device_async_resource_ref alternate_upstream)
+    : primary_upstream_{primary_upstream}, alternate_upstream_{alternate_upstream}
+  {
+  }
+
+  failure_alternate_resource_adaptor()                                          = delete;
+  ~failure_alternate_resource_adaptor() override                                = default;
+  failure_alternate_resource_adaptor(failure_alternate_resource_adaptor const&) = delete;
+  failure_alternate_resource_adaptor& operator=(failure_alternate_resource_adaptor const&) = delete;
+  failure_alternate_resource_adaptor(failure_alternate_resource_adaptor&&) noexcept =
+    default;  ///< @default_move_constructor
+  failure_alternate_resource_adaptor& operator=(failure_alternate_resource_adaptor&&) noexcept =
+    default;  ///< @default_move_assignment{failure_alternate_resource_adaptor}
+
+  /**
+   * @briefreturn{rmm::device_async_resource_ref to the upstream resource}
+   */
+  [[nodiscard]] rmm::device_async_resource_ref get_upstream_resource() const noexcept
+  {
+    return primary_upstream_;
+  }
+
+  /**
+   * @briefreturn{rmm::device_async_resource_ref to the alternate upstream resource}
+   */
+  [[nodiscard]] rmm::device_async_resource_ref get_alternate_upstream_resource() const noexcept
+  {
+    return alternate_upstream_;
+  }
+
+ private:
+  /**
+   * @brief Allocates memory of size at least `bytes` using the upstream
+   * resource.
+   *
+   * @throws any exceptions thrown from the upstream resources, only `exception_type`
+   * thrown by the primary upstream is caught.
+   *
+   * @param bytes The size, in bytes, of the allocation
+   * @param stream Stream on which to perform the allocation
+   * @return void* Pointer to the newly allocated memory
+   */
+  void* do_allocate(std::size_t bytes, cuda_stream_view stream) override
+  {
+    void* ret{};
+    try {
+      ret = primary_upstream_.allocate_async(bytes, stream);
+    } catch (exception_type const& e) {
+      ret = alternate_upstream_.allocate_async(bytes, stream);
+      std::lock_guard<std::mutex> lock(mtx_);
+      alternate_allocations_.insert(ret);
+    }
+    return ret;
+  }
+
+  /**
+   * @brief Free allocation of size `bytes` pointed to by `ptr`
+   *
+   * @param ptr Pointer to be deallocated
+   * @param bytes Size of the allocation
+   * @param stream Stream on which to perform the deallocation
+   */
+  void do_deallocate(void* ptr, std::size_t bytes, cuda_stream_view stream) override
+  {
+    std::size_t count{0};
+    {
+      std::lock_guard<std::mutex> lock(mtx_);
+      count = alternate_allocations_.erase(ptr);
+    }
+    if (count > 0) {
+      alternate_upstream_.deallocate_async(ptr, bytes, stream);
+    } else {
+      primary_upstream_.deallocate_async(ptr, bytes, stream);
+    }
+  }
+
+  /**
+   * @brief Compare the resource to another.
+   *
+   * @param other The other resource to compare to
+   * @return true If the two resources are equivalent
+   * @return false If the two resources are not equal
+   */
+  [[nodiscard]] bool do_is_equal(device_memory_resource const& other) const noexcept override
+  {
+    if (this == &other) { return true; }
+    auto cast = dynamic_cast<failure_alternate_resource_adaptor const*>(&other);
+    if (cast == nullptr) { return false; }
+    return get_upstream_resource() == cast->get_upstream_resource() &&
+           get_alternate_upstream_resource() == cast->get_alternate_upstream_resource();
+  }
+
+  device_async_resource_ref primary_upstream_;
+  device_async_resource_ref alternate_upstream_;
+  std::unordered_set<void*> alternate_allocations_;
+  mutable std::mutex mtx_;
+};
+
+/** @} */  // end of group
+}  // namespace mr
+}  // namespace RMM_NAMESPACE

python/rmm/rmm/_lib/memory_resource.pxd

@@ -97,6 +97,12 @@ cdef class TrackingResourceAdaptor(UpstreamResourceAdaptor):
 cdef class FailureCallbackResourceAdaptor(UpstreamResourceAdaptor):
     cdef object _callback
 
+cdef class FailureAlternateResourceAdaptor(UpstreamResourceAdaptor):
+    cdef readonly DeviceMemoryResource alternate_upstream_mr
+
+    cpdef DeviceMemoryResource get_alternate_upstream(self)
+
+
 cdef class PrefetchResourceAdaptor(UpstreamResourceAdaptor):
     pass
 

python/rmm/rmm/_lib/memory_resource.pyx

@@ -32,7 +32,9 @@ from libcpp.string cimport string
 from cuda.cudart import cudaError_t
 
 from rmm._cuda.gpu import CUDARuntimeError, getDevice, setDevice
+
 from rmm._cuda.stream cimport Stream
+
 from rmm._cuda.stream import DEFAULT_STREAM
 
 from rmm._lib.cuda_stream_view cimport cuda_stream_view
@@ -44,6 +46,7 @@ from rmm._lib.per_device_resource cimport (
     cuda_device_id,
     set_per_device_resource as cpp_set_per_device_resource,
 )
+
 from rmm.statistics import Statistics
 
 # Transparent handle of a C++ exception
@@ -84,6 +87,10 @@ cdef extern from *:
 
 # NOTE: Keep extern declarations in .pyx file as much as possible to avoid
 # leaking dependencies when importing RMM Cython .pxd files
+
+cdef extern from "rmm/error.hpp" namespace "rmm" nogil:
+    cdef cppclass out_of_memory
+
 cdef extern from "rmm/mr/device/cuda_memory_resource.hpp" \
         namespace "rmm::mr" nogil:
     cdef cppclass cuda_memory_resource(device_memory_resource):
@@ -125,7 +132,6 @@ cdef extern from "rmm/mr/device/cuda_async_memory_resource.hpp" \
         win32
         win32_kmt
 
-
 cdef extern from "rmm/mr/device/pool_memory_resource.hpp" \
         namespace "rmm::mr" nogil:
     cdef cppclass pool_memory_resource[Upstream](device_memory_resource):
@@ -229,6 +235,19 @@ cdef extern from "rmm/mr/device/failure_callback_resource_adaptor.hpp" \
             void* callback_arg
         ) except +
 
+cdef extern from "rmm/mr/device/failure_alternate_resource_adaptor.hpp" \
+        namespace "rmm::mr" nogil:
+    cdef cppclass failure_alternate_resource_adaptor[ExceptionType](
+        device_memory_resource
+    ):
+        # Notice, `failure_alternate_resource_adaptor` takes `device_async_resource_ref`
+        # as upstream arguments but we define them here as `device_memory_resource*` and
+        # rely on implicit type conversion.
+        failure_alternate_resource_adaptor(
+            device_memory_resource* upstream_mr,
+            device_memory_resource* alternate_upstream_mr,
+        ) except +
+
 cdef extern from "rmm/mr/device/prefetch_resource_adaptor.hpp" \
         namespace "rmm::mr" nogil:
     cdef cppclass prefetch_resource_adaptor[Upstream](device_memory_resource):
@@ -279,7 +298,6 @@ cdef class UpstreamResourceAdaptor(DeviceMemoryResource):
     """
 
     def __cinit__(self, DeviceMemoryResource upstream_mr, *args, **kwargs):
-
         if (upstream_mr is None):
             raise Exception("Argument `upstream_mr` must not be None")
 
@@ -1039,6 +1057,46 @@ cdef class FailureCallbackResourceAdaptor(UpstreamResourceAdaptor):
         """
         pass
 
+
+cdef class FailureAlternateResourceAdaptor(UpstreamResourceAdaptor):
+
+    def __cinit__(
+        self,
+        DeviceMemoryResource upstream_mr,
+        DeviceMemoryResource alternate_upstream_mr,
+    ):
+        if (alternate_upstream_mr is None):
+            raise Exception("Argument `alternate_upstream_mr` must not be None")
+        self.alternate_upstream_mr = alternate_upstream_mr
+
+        self.c_obj.reset(
+            new failure_alternate_resource_adaptor[out_of_memory](
+                upstream_mr.get_mr(),
+                alternate_upstream_mr.get_mr(),
+            )
+        )
+
+    def __init__(
+        self,
+        DeviceMemoryResource upstream_mr,
+        DeviceMemoryResource alternate_upstream_mr,
+    ):
+        """
+        A memory resource that uses an alternate resource when memory allocation fails.
+
+        Parameters
+        ----------
+        upstream : DeviceMemoryResource
+            The primary resource used for allocating/deallocating device memory
+        alternate_upstream : DeviceMemoryResource
+            The alternate resource used when the primary fails to allocate
+        """
+        pass
+
+    cpdef DeviceMemoryResource get_alternate_upstream(self):
+        return self.alternate_upstream_mr
+
+
 cdef class PrefetchResourceAdaptor(UpstreamResourceAdaptor):
 
     def __cinit__(

python/rmm/rmm/mr.py

@@ -17,6 +17,7 @@
     CudaAsyncMemoryResource,
     CudaMemoryResource,
     DeviceMemoryResource,
+    FailureAlternateResourceAdaptor,
     FailureCallbackResourceAdaptor,
     FixedSizeMemoryResource,
     LimitingResourceAdaptor,
@@ -61,6 +62,7 @@
     "SystemMemoryResource",
     "TrackingResourceAdaptor",
     "FailureCallbackResourceAdaptor",
+    "FailureAlternateResourceAdaptor",
     "UpstreamResourceAdaptor",
     "_flush_logs",
     "_initialize",

python/rmm/rmm/tests/test_rmm.py

@@ -787,6 +787,57 @@ def callback(nbytes: int) -> bool:
     assert retried[0]
 
 
+def test_failure_alternate_resource_adaptor():
+    base = rmm.mr.CudaMemoryResource()
+
+    def alloc_cb(size, stream, *, track: list[int], limit: int):
+        if size > limit:
+            raise MemoryError()
+        ret = base.allocate(size, stream)
+        track.append(ret)
+        return ret
+
+    def dealloc_cb(ptr, size, stream, *, track: list[int]):
+        track.append(ptr)
+        return base.deallocate(ptr, size, stream)
+
+    main_track = []
+    main_mr = rmm.mr.CallbackMemoryResource(
+        functools.partial(alloc_cb, track=main_track, limit=200),
+        functools.partial(dealloc_cb, track=main_track),
+    )
+    alternate_track = []
+    alternate_mr = rmm.mr.CallbackMemoryResource(
+        functools.partial(alloc_cb, track=alternate_track, limit=1000),
+        functools.partial(dealloc_cb, track=alternate_track),
+    )
+    mr = rmm.mr.FailureAlternateResourceAdaptor(main_mr, alternate_mr)
+    assert main_mr is mr.get_upstream()
+    assert alternate_mr is mr.get_alternate_upstream()
+
+    # Delete the upstream memory resources here to check that they are
+    # kept alive by `mr`
+    del main_mr
+    del alternate_mr
+
+    # Buffer size within the limit of `main_mr`
+    rmm.DeviceBuffer(size=100, mr=mr)
+    # we expect an alloc and a dealloc of the same buffer in
+    # `main_track` and an empty `alternate_track`
+    assert len(main_track) == 2
+    assert main_track[0] == main_track[1]
+    assert len(alternate_track) == 0
+
+    # Buffer size outside the limit of `main_mr`
+    rmm.DeviceBuffer(size=500, mr=mr)
+    # we expect an alloc and a dealloc of the same buffer in
+    # `alternate_track` and an unchanged `main_mr`
+    assert len(main_track) == 2
+    assert main_track[0] == main_track[1]
+    assert len(alternate_track) == 2
+    assert alternate_track[0] == alternate_track[1]
+
+
 @pytest.mark.parametrize("managed", [True, False])
 def test_prefetch_resource_adaptor(managed):
     if managed:

tests/CMakeLists.txt

@@ -145,9 +145,12 @@ ConfigureTest(STATISTICS_TEST mr/device/statistics_mr_tests.cpp GPUS 1 PERCENT 1
 # tracking adaptor tests
 ConfigureTest(TRACKING_TEST mr/device/tracking_mr_tests.cpp GPUS 1 PERCENT 100)
 
-# out-of-memory callback adaptor tests
+# failure callback adaptor tests
 ConfigureTest(FAILURE_CALLBACK_TEST mr/device/failure_callback_mr_tests.cpp)
 
+# failure alternate adaptor tests
+ConfigureTest(FAILURE_ALTERNATE_TEST mr/device/failure_alternate_mr_tests.cpp)
+
 # prefetch adaptor tests
 ConfigureTest(PREFETCH_ADAPTOR_TEST mr/device/prefetch_resource_adaptor_tests.cpp)