Skip to content

Latest commit

 

History

History
971 lines (725 loc) · 40.4 KB

README.md

File metadata and controls

971 lines (725 loc) · 40.4 KB
Raw

 RMM: RAPIDS Memory Manager

NOTE: For the latest stable README.md ensure you are on the main branch.

Resources

Overview

Achieving optimal performance in GPU-centric workflows frequently requires customizing how host and device memory are allocated. For example, using "pinned" host memory for asynchronous host <-> device memory transfers, or using a device memory pool sub-allocator to reduce the cost of dynamic device memory allocation.

The goal of the RAPIDS Memory Manager (RMM) is to provide:

  • A common interface that allows customizing device and host memory allocation
  • A collection of implementations of the interface
  • A collection of data structures that use the interface for memory allocation

For information on the interface RMM provides and how to use RMM in your C++ code, see below.

For a walkthrough about the design of the RAPIDS Memory Manager, read Fast, Flexible Allocation for NVIDIA CUDA with RAPIDS Memory Manager on the NVIDIA Developer Blog.

Installation

Conda

RMM can be installed with conda. You can get a minimal conda installation with miniforge.

Install RMM with:

conda install -c rapidsai -c conda-forge -c nvidia rmm cuda-version=12.0

We also provide nightly conda packages built from the HEAD of our latest development branch.

Note: RMM is supported only on Linux, and only tested with Python versions 3.10, 3.11, and 3.12.

Note: The RMM package from conda requires building with GCC 9 or later. Otherwise, your application may fail to build.

See the Get RAPIDS version picker for more OS and version info.

Building from Source

Get RMM Dependencies

Compiler requirements:

  • gcc version 9.3+
  • nvcc version 11.4+
  • cmake version 3.26.4+

CUDA/GPU requirements:

GPU Support:

  • RMM is tested and supported only on Volta architecture and newer (Compute Capability 7.0+). It may work on earlier architectures.

Python requirements:

  • rapids-build-backend (available from PyPI or the rapidsai conda channel)
  • scikit-build-core
  • cuda-python
  • cython

For more details, see pyproject.toml

Script to build RMM from source

To install RMM from source, ensure the dependencies are met and follow the steps below:

  • Clone the repository and submodules
$ git clone --recurse-submodules https://github.com/rapidsai/rmm.git
$ cd rmm
  • Create the conda development environment rmm_dev
# create the conda environment (assuming in base `rmm` directory)
$ conda env create --name rmm_dev --file conda/environments/all_cuda-118_arch-x86_64.yaml
# activate the environment
$ conda activate rmm_dev
  • Build and install librmm using cmake & make. CMake depends on the nvcc executable being on your path or defined in CUDACXX environment variable.
$ mkdir build                                       # make a build directory
$ cd build                                          # enter the build directory
$ cmake .. -DCMAKE_INSTALL_PREFIX=/install/path     # configure cmake ... use $CONDA_PREFIX if you're using Anaconda
$ make -j                                           # compile the library librmm.so ... '-j' will start a parallel job using the number of physical cores available on your system
$ make install                                      # install the library librmm.so to '/install/path'
  • Building and installing librmm and rmm using build.sh. Build.sh creates build dir at root of git repository. build.sh depends on the nvcc executable being on your path or defined in CUDACXX environment variable.
$ ./build.sh -h                                     # Display help and exit
$ ./build.sh -n librmm                              # Build librmm without installing
$ ./build.sh -n rmm                                 # Build rmm without installing
$ ./build.sh -n librmm rmm                          # Build librmm and rmm without installing
$ ./build.sh librmm rmm                             # Build and install librmm and rmm
  • To run tests (Optional):
$ cd build (if you are not already in build directory)
$ make test
  • Build, install, and test the rmm python package, in the python folder:
# In the root rmm directory
$ python -m pip install -e ./python/rmm
$ pytest -v

Done! You are ready to develop for the RMM OSS project.

Caching third-party dependencies

RMM uses CPM.cmake to handle third-party dependencies like spdlog, Thrust, GoogleTest, GoogleBenchmark. In general you won't have to worry about it. If CMake finds an appropriate version on your system, it uses it (you can help it along by setting CMAKE_PREFIX_PATH to point to the installed location). Otherwise those dependencies will be downloaded as part of the build.

If you frequently start new builds from scratch, consider setting the environment variable CPM_SOURCE_CACHE to an external download directory to avoid repeated downloads of the third-party dependencies.

Using RMM in a downstream CMake project

The installed RMM library provides a set of config files that makes it easy to integrate RMM into your own CMake project. In your CMakeLists.txt, just add

find_package(rmm [VERSION])
# ...
target_link_libraries(<your-target> (PRIVATE|PUBLIC|INTERFACE) rmm::rmm)

Since RMM is a header-only library, this does not actually link RMM, but it makes the headers available and pulls in transitive dependencies. If RMM is not installed in a default location, use CMAKE_PREFIX_PATH or rmm_ROOT to point to its location.

One of RMM's dependencies is the Thrust library, so the above automatically pulls in Thrust by means of a dependency on the rmm::Thrust target. By default it uses the standard configuration of Thrust. If you want to customize it, you can set the variables THRUST_HOST_SYSTEM and THRUST_DEVICE_SYSTEM; see Thrust's CMake documentation.

Using CPM to manage RMM

RMM uses CPM.cmake to manage its dependencies, including CCCL, and you can use CPM for your project's dependency on RMM.

There is an issue with using CPM's single-argument compact syntax for RMM/CCCL as it transitively marks targets as SYSTEM dependencies. This causes the CCCL headers pulled in through CPM to be of lower priority to the preprocessor than the (potentially outdated) CCCL headers provided by the CUDA SDK. To avoid this issue, use CPM's multi-argument syntax instead:

CPMAddPackage(NAME rmm [VERSION]
              GITHUB_REPOSITORY rapidsai/rmm
              SYSTEM Off)
# ...
target_link_libraries(<your-target> (PRIVATE|PUBLIC|INTERFACE) rmm::rmm)

Using RMM in C++

The first goal of RMM is to provide a common interface for device and host memory allocation. This allows both users and implementers of custom allocation logic to program to a single interface.

To this end, RMM defines two abstract interface classes:

These classes are based on the std::pmr::memory_resource interface class introduced in C++17 for polymorphic memory allocation.

device_memory_resource

rmm::mr::device_memory_resource is the base class that defines the interface for allocating and freeing device memory.

It has two key functions:

  1. void* device_memory_resource::allocate(std::size_t bytes, cuda_stream_view s)

    • Returns a pointer to an allocation of at least bytes bytes.

  2. void device_memory_resource::deallocate(void* p, std::size_t bytes, cuda_stream_view s)

    • Reclaims a previous allocation of size bytes pointed to by p.
    • p must have been returned by a previous call to allocate(bytes), otherwise behavior is undefined

It is up to a derived class to provide implementations of these functions. See available resources for example device_memory_resource derived classes.

Unlike std::pmr::memory_resource, rmm::mr::device_memory_resource does not allow specifying an alignment argument. All allocations are required to be aligned to at least 256B. Furthermore, device_memory_resource adds an additional cuda_stream_view argument to allow specifying the stream on which to perform the (de)allocation.

Stream-ordered Memory Allocation

rmm::mr::device_memory_resource is a base class that provides stream-ordered memory allocation. This allows optimizations such as re-using memory deallocated on the same stream without the overhead of synchronization.

A call to device_memory_resource::allocate(bytes, stream_a) returns a pointer that is valid to use on stream_a. Using the memory on a different stream (say stream_b) is Undefined Behavior unless the two streams are first synchronized, for example by using cudaStreamSynchronize(stream_a) or by recording a CUDA event on stream_a and then calling cudaStreamWaitEvent(stream_b, event).

The stream specified to device_memory_resource::deallocate should be a stream on which it is valid to use the deallocated memory immediately for another allocation. Typically this is the stream on which the allocation was last used before the call to deallocate. The passed stream may be used internally by a device_memory_resource for managing available memory with minimal synchronization, and it may also be synchronized at a later time, for example using a call to cudaStreamSynchronize().

For this reason, it is Undefined Behavior to destroy a CUDA stream that is passed to device_memory_resource::deallocate. If the stream on which the allocation was last used has been destroyed before calling deallocate or it is known that it will be destroyed, it is likely better to synchronize the stream (before destroying it) and then pass a different stream to deallocate (e.g. the default stream).

Note that device memory data structures such as rmm::device_buffer and rmm::device_uvector follow these stream-ordered memory allocation semantics and rules.

For further information about stream-ordered memory allocation semantics, read Using the NVIDIA CUDA Stream-Ordered Memory Allocator on the NVIDIA Developer Blog.

Available Device Resources

RMM provides several device_memory_resource derived classes to satisfy various user requirements. For more detailed information about these resources, see their respective documentation.

cuda_memory_resource

Allocates and frees device memory using cudaMalloc and cudaFree.

managed_memory_resource

Allocates and frees device memory using cudaMallocManaged and cudaFree.

Note that managed_memory_resource cannot be used with NVIDIA Virtual GPU Software (vGPU, for use with virtual machines or hypervisors) because NVIDIA CUDA Unified Memory is not supported by NVIDIA vGPU.

pool_memory_resource

A coalescing, best-fit pool sub-allocator.

fixed_size_memory_resource

A memory resource that can only allocate a single fixed size. Average allocation and deallocation cost is constant.

binning_memory_resource

Configurable to use multiple upstream memory resources for allocations that fall within different bin sizes. Often configured with multiple bins backed by fixed_size_memory_resources and a single pool_memory_resource for allocations larger than the largest bin size.

Default Resources and Per-device Resources

RMM users commonly need to configure a device_memory_resource object to use for all allocations where another resource has not explicitly been provided. A common example is configuring a pool_memory_resource to use for all allocations to get fast dynamic allocation.

To enable this use case, RMM provides the concept of a "default" device_memory_resource. This resource is used when another is not explicitly provided.

Accessing and modifying the default resource is done through two functions:

  • device_memory_resource* get_current_device_resource()

    • Returns a pointer to the default resource for the current CUDA device.
    • The initial default memory resource is an instance of cuda_memory_resource.
    • This function is thread safe with respect to concurrent calls to it and set_current_device_resource().
    • For more explicit control, you can use get_per_device_resource(), which takes a device ID.

  • device_memory_resource* set_current_device_resource(device_memory_resource* new_mr)

    • Updates the default memory resource pointer for the current CUDA device to new_mr
    • Returns the previous default resource pointer
    • If new_mr is nullptr, then resets the default resource to cuda_memory_resource
    • This function is thread safe with respect to concurrent calls to it and get_current_device_resource()
    • For more explicit control, you can use set_per_device_resource(), which takes a device ID.

Example

rmm::mr::cuda_memory_resource cuda_mr;
// Construct a resource that uses a coalescing best-fit pool allocator
// With the pool initially half of available device memory
auto initial_size = rmm::percent_of_free_device_memory(50);
rmm::mr::pool_memory_resource<rmm::mr::cuda_memory_resource> pool_mr{&cuda_mr, initial_size};
rmm::mr::set_current_device_resource(&pool_mr); // Updates the current device resource pointer to `pool_mr`
rmm::mr::device_memory_resource* mr = rmm::mr::get_current_device_resource(); // Points to `pool_mr`

Multiple Devices

A device_memory_resource should only be used when the active CUDA device is the same device that was active when the device_memory_resource was created. Otherwise behavior is undefined.

If a device_memory_resource is used with a stream associated with a different CUDA device than the device for which the memory resource was created, behavior is undefined.

Creating a device_memory_resource for each device requires care to set the current device before creating each resource, and to maintain the lifetime of the resources as long as they are set as per-device resources. Here is an example loop that creates unique_ptrs to pool_memory_resource objects for each device and sets them as the per-device resource for that device.

using pool_mr = rmm::mr::pool_memory_resource<rmm::mr::cuda_memory_resource>;
std::vector<unique_ptr<pool_mr>> per_device_pools;
for(int i = 0; i < N; ++i) {
  cudaSetDevice(i); // set device i before creating MR
  // Use a vector of unique_ptr to maintain the lifetime of the MRs
  // Note: for brevity, omitting creation of upstream and computing initial_size
  per_device_pools.push_back(std::make_unique<pool_mr>(upstream, initial_size));
  // Set the per-device resource for device i
  set_per_device_resource(cuda_device_id{i}, &per_device_pools.back());
}

Note that the CUDA device that is current when creating a device_memory_resource must also be current any time that device_memory_resource is used to deallocate memory, including in a destructor. The RAII class rmm::device_buffer and classes that use it as a backing store (rmm::device_scalar and rmm::device_uvector) handle this by storing the active device when the constructor is called, and then ensuring that the stored device is active whenever an allocation or deallocation is performed (including in the destructor). The user must therefore only ensure that the device active during creation of an rmm::device_buffer matches the active device of the memory resource being used.

Here is an incorrect example that creates a memory resource on device zero and then uses it to allocate a device_buffer on device one:

{
  RMM_CUDA_TRY(cudaSetDevice(0));
  auto mr = rmm::mr::cuda_memory_resource{};
  {
    RMM_CUDA_TRY(cudaSetDevice(1));
    // Invalid, current device is 1, but MR is only valid for device 0
    rmm::device_buffer buf(16, rmm::cuda_stream_default, &mr);
  }
}

A correct example creates the device buffer with device zero active. After that it is safe to switch devices and let the buffer go out of scope and destruct with a different device active. For example, this code is correct:

{
  RMM_CUDA_TRY(cudaSetDevice(0));
  auto mr = rmm::mr::cuda_memory_resource{};
  rmm::device_buffer buf(16, rmm::cuda_stream_default, &mr);
  RMM_CUDA_TRY(cudaSetDevice(1));
  ...
  // No need to switch back to device 0 before ~buf runs
}

Use of rmm::device_vector with multiple devices

rmm:device_vector uses an rmm::mr::thrust_allocator to enable thrust::device_vector to allocate and deallocate memory using RMM. As such, the usual rules for usage of the backing memory resource apply: the active device must match the active device at resource construction time. To facilitate use in an RAII setting, rmm::mr::thrust_allocator records the active device at construction time and ensures that device is active whenever it allocates or deallocates memory. Usage of rmm::device_vector with multiple devices is therefore the same as rmm::device_buffer. One must create device_vectors with the correct device active, but it is safe to destroy them with a different active device.

For example, recapitulating the previous example using rmm::device_vector:

{
  RMM_CUDA_TRY(cudaSetDevice(0));
  auto mr = rmm::mr::cuda_memory_resource{};
  rmm::device_vector<int> vec(16, rmm::mr::thrust_allocator<int>(rmm::cuda_stream_default, &mr));
  RMM_CUDA_TRY(cudaSetDevice(1));
  ...
  // No need to switch back to device 0 before ~vec runs
}

Note

Although allocation and deallocation in the thrust_allocator run with the correct active device, modification of rmm::device_vector might necessitate a kernel launch, and this must run with the correct device active. For example, .resize() might both allocate and launch a kernel to initialize new elements: the user must arrange for this kernel launch to occur with the correct device for the memory resource active.

cuda_stream_view and cuda_stream

rmm::cuda_stream_view is a simple non-owning wrapper around a CUDA cudaStream_t. This wrapper's purpose is to provide strong type safety for stream types. (cudaStream_t is an alias for a pointer, which can lead to ambiguity in APIs when it is assigned 0.) All RMM stream-ordered APIs take a rmm::cuda_stream_view argument.

rmm::cuda_stream is a simple owning wrapper around a CUDA cudaStream_t. This class provides RAII semantics (constructor creates the CUDA stream, destructor destroys it). An rmm::cuda_stream can never represent the CUDA default stream or per-thread default stream; it only ever represents a single non-default stream. rmm::cuda_stream cannot be copied, but can be moved.

cuda_stream_pool

rmm::cuda_stream_pool provides fast access to a pool of CUDA streams. This class can be used to create a set of cuda_stream objects whose lifetime is equal to the cuda_stream_pool. Using the stream pool can be faster than creating the streams on the fly. The size of the pool is configurable. Depending on this size, multiple calls to cuda_stream_pool::get_stream() may return instances of rmm::cuda_stream_view that represent identical CUDA streams.

Thread Safety

All current device memory resources are thread safe unless documented otherwise. More specifically, calls to memory resource allocate() and deallocate() methods are safe with respect to calls to either of these functions from other threads. They are not thread safe with respect to construction and destruction of the memory resource object.

Note that a class thread_safe_resource_adapter is provided which can be used to adapt a memory resource that is not thread safe to be thread safe (as described above). This adapter is not needed with any current RMM device memory resources.

Allocators

C++ interfaces commonly allow customizable memory allocation through an Allocator object. RMM provides several Allocator and Allocator-like classes.

polymorphic_allocator

A stream-ordered allocator similar to std::pmr::polymorphic_allocator. Unlike the standard C++ Allocator interface, the allocate and deallocate functions take a cuda_stream_view indicating the stream on which the (de)allocation occurs.

stream_allocator_adaptor

stream_allocator_adaptor can be used to adapt a stream-ordered allocator to present a standard Allocator interface to consumers that may not be designed to work with a stream-ordered interface.

Example:

rmm::cuda_stream stream;
rmm::mr::polymorphic_allocator<int> stream_alloc;

// Constructs an adaptor that forwards all (de)allocations to `stream_alloc` on `stream`.
auto adapted = rmm::mr::stream_allocator_adaptor(stream_alloc, stream);

// Allocates 100 bytes using `stream_alloc` on `stream`
auto p = adapted.allocate(100);
...
// Deallocates using `stream_alloc` on `stream`
adapted.deallocate(p,100);

thrust_allocator

thrust_allocator is a device memory allocator that uses the strongly typed thrust::device_ptr, making it usable with containers like thrust::device_vector.

See below for more information on using RMM with Thrust.

Device Data Structures

device_buffer

An untyped, uninitialized RAII class for stream ordered device memory allocation.

Example

cuda_stream_view s{...};
// Allocates at least 100 bytes on stream `s` using the *default* resource
rmm::device_buffer b{100,s};
void* p = b.data();                   // Raw, untyped pointer to underlying device memory

kernel<<<..., s.value()>>>(b.data()); // `b` is only safe to use on `s`

rmm::mr::device_memory_resource * mr = new my_custom_resource{...};
// Allocates at least 100 bytes on stream `s` using the resource `mr`
rmm::device_buffer b2{100, s, mr};

device_uvector<T>

A typed, uninitialized RAII class for allocation of a contiguous set of elements in device memory. Similar to a thrust::device_vector, but as an optimization, does not default initialize the contained elements. This optimization restricts the types T to trivially copyable types.

Example

cuda_stream_view s{...};
// Allocates uninitialized storage for 100 `int32_t` elements on stream `s` using the
// default resource
rmm::device_uvector<int32_t> v(100, s);
// Initializes the elements to 0
thrust::uninitialized_fill(thrust::cuda::par.on(s.value()), v.begin(), v.end(), int32_t{0});

rmm::mr::device_memory_resource * mr = new my_custom_resource{...};
// Allocates uninitialized storage for 100 `int32_t` elements on stream `s` using the resource `mr`
rmm::device_uvector<int32_t> v2{100, s, mr};

device_scalar

A typed, RAII class for allocation of a single element in device memory. This is similar to a device_uvector with a single element, but provides convenience functions like modifying the value in device memory from the host, or retrieving the value from device to host.

Example

cuda_stream_view s{...};
// Allocates uninitialized storage for a single `int32_t` in device memory
rmm::device_scalar<int32_t> a{s};
a.set_value(42, s); // Updates the value in device memory to `42` on stream `s`

kernel<<<...,s.value()>>>(a.data()); // Pass raw pointer to underlying element in device memory

int32_t v = a.value(s); // Retrieves the value from device to host on stream `s`

host_memory_resource

rmm::mr::host_memory_resource is the base class that defines the interface for allocating and freeing host memory.

Similar to device_memory_resource, it has two key functions for (de)allocation:

  1. void* host_memory_resource::allocate(std::size_t bytes, std::size_t alignment)

    • Returns a pointer to an allocation of at least bytes bytes aligned to the specified alignment

  2. void host_memory_resource::deallocate(void* p, std::size_t bytes, std::size_t alignment)

    • Reclaims a previous allocation of size bytes pointed to by p.

Unlike device_memory_resource, the host_memory_resource interface and behavior is identical to std::pmr::memory_resource.

Available Host Resources

new_delete_resource

Uses the global operator new and operator delete to allocate host memory.

pinned_memory_resource

Allocates "pinned" host memory using cuda(Malloc/Free)Host.

Host Data Structures

RMM does not currently provide any data structures that interface with host_memory_resource. In the future, RMM will provide a similar host-side structure like device_buffer and an allocator that can be used with STL containers.

Using RMM with Thrust

RAPIDS and other CUDA libraries make heavy use of Thrust. Thrust uses CUDA device memory in two situations:

  1. As the backing store for thrust::device_vector, and
  2. As temporary storage inside some algorithms, such as thrust::sort.

RMM provides rmm::mr::thrust_allocator as a conforming Thrust allocator that uses device_memory_resources.

Thrust Algorithms

To instruct a Thrust algorithm to use rmm::mr::thrust_allocator to allocate temporary storage, you can use the custom Thrust CUDA device execution policy: rmm::exec_policy(stream).

thrust::sort(rmm::exec_policy(stream, ...);

The first stream argument is the stream to use for rmm::mr::thrust_allocator. The second stream argument is what should be used to execute the Thrust algorithm. These two arguments must be identical.

Logging

RMM includes two forms of logging. Memory event logging and debug logging.

Memory Event Logging and logging_resource_adaptor

Memory event logging writes details of every allocation or deallocation to a CSV (comma-separated value) file. In C++, Memory Event Logging is enabled by using the logging_resource_adaptor as a wrapper around any other device_memory_resource object.

Each row in the log represents either an allocation or a deallocation. The columns of the file are "Thread, Time, Action, Pointer, Size, Stream".

The CSV output files of the logging_resource_adaptor can be used as input to REPLAY_BENCHMARK, which is available when building RMM from source, in the gbenchmarks folder in the build directory. This log replayer can be useful for profiling and debugging allocator issues.

The following C++ example creates a logging version of a cuda_memory_resource that outputs the log to the file "logs/test1.csv".

std::string filename{"logs/test1.csv"};
rmm::mr::cuda_memory_resource upstream;
rmm::mr::logging_resource_adaptor<rmm::mr::cuda_memory_resource> log_mr{&upstream, filename};

If a file name is not specified, the environment variable RMM_LOG_FILE is queried for the file name. If RMM_LOG_FILE is not set, then an exception is thrown by the logging_resource_adaptor constructor.

In Python, memory event logging is enabled when the logging parameter of rmm.reinitialize() is set to True. The log file name can be set using the log_file_name parameter. See help(rmm.reinitialize) for full details.

Debug Logging

RMM includes a debug logger which can be enabled to log trace and debug information to a file. This information can show when errors occur, when additional memory is allocated from upstream resources, etc. The default log file is rmm_log.txt in the current working directory, but the environment variable RMM_DEBUG_LOG_FILE can be set to specify the path and file name.

There is a CMake configuration variable RMM_LOGGING_LEVEL, which can be set to enable compilation of more detailed logging. The default is INFO. Available levels are TRACE, DEBUG, INFO, WARN, ERROR, CRITICAL and OFF.

The log relies on the spdlog library.

Note that to see logging below the INFO level, the application must also set the logging level at run time. C++ applications must must call rmm::logger().set_level(), for example to enable all levels of logging down to TRACE, call rmm::logger().set_level(spdlog::level::trace) (and compile librmm with -DRMM_LOGGING_LEVEL=TRACE). Python applications must call rmm.set_logging_level(), for example to enable all levels of logging down to TRACE, call rmm.set_logging_level("trace") (and compile the RMM Python module with -DRMM_LOGGING_LEVEL=TRACE).

Note that debug logging is different from the CSV memory allocation logging provided by rmm::mr::logging_resource_adapter. The latter is for logging a history of allocation / deallocation actions which can be useful for replay with RMM's replay benchmark.

RMM and CUDA Memory Bounds Checking

Memory allocations taken from a memory resource that allocates a pool of memory (such as pool_memory_resource and arena_memory_resource) are part of the same low-level CUDA memory allocation. Therefore, out-of-bounds or misaligned accesses to these allocations are not likely to be detected by CUDA tools such as CUDA Compute Sanitizer memcheck.

Exceptions to this are cuda_memory_resource, which wraps cudaMalloc, and cuda_async_memory_resource, which uses cudaMallocAsync with CUDA's built-in memory pool functionality (CUDA 11.2 or later required). Illegal memory accesses to memory allocated by these resources are detectable with Compute Sanitizer Memcheck.

It may be possible in the future to add support for memory bounds checking with other memory resources using NVTX APIs.

Using RMM in Python

There are two ways to use RMM in Python code:

  1. Using the rmm.DeviceBuffer API to explicitly create and manage device memory allocations
  2. Transparently via external libraries such as CuPy and Numba

RMM provides a MemoryResource abstraction to control how device memory is allocated in both the above uses.

DeviceBuffer

A DeviceBuffer represents an untyped, uninitialized device memory allocation. DeviceBuffers can be created by providing the size of the allocation in bytes:

>>> import rmm
>>> buf = rmm.DeviceBuffer(size=100)

The size of the allocation and the memory address associated with it can be accessed via the .size and .ptr attributes respectively:

>>> buf.size
100
>>> buf.ptr
140202544726016

DeviceBuffers can also be created by copying data from host memory:

>>> import rmm
>>> import numpy as np
>>> a = np.array([1, 2, 3], dtype='float64')
>>> buf = rmm.DeviceBuffer.to_device(a.tobytes())
>>> buf.size
24

Conversely, the data underlying a DeviceBuffer can be copied to the host:

>>> np.frombuffer(buf.tobytes())
array([1., 2., 3.])

MemoryResource objects

MemoryResource objects are used to configure how device memory allocations are made by RMM.

By default if a MemoryResource is not set explicitly, RMM uses the CudaMemoryResource, which uses cudaMalloc for allocating device memory.

rmm.reinitialize() provides an easy way to initialize RMM with specific memory resource options across multiple devices. See help(rmm.reinitialize) for full details.

For lower-level control, the rmm.mr.set_current_device_resource() function can be used to set a different MemoryResource for the current CUDA device. For example, enabling the ManagedMemoryResource tells RMM to use cudaMallocManaged instead of cudaMalloc for allocating memory:

>>> import rmm
>>> rmm.mr.set_current_device_resource(rmm.mr.ManagedMemoryResource())

⚠️ The default resource must be set for any device before allocating any device memory on that device. Setting or changing the resource after device allocations have been made can lead to unexpected behaviour or crashes. See Multiple Devices

As another example, PoolMemoryResource allows you to allocate a large "pool" of device memory up-front. Subsequent allocations will draw from this pool of already allocated memory. The example below shows how to construct a PoolMemoryResource with an initial size of 1 GiB and a maximum size of 4 GiB. The pool uses CudaMemoryResource as its underlying ("upstream") memory resource:

>>> import rmm
>>> pool = rmm.mr.PoolMemoryResource(
...     rmm.mr.CudaMemoryResource(),
...     initial_pool_size="1GiB", # equivalent to initial_pool_size=2**30
...     maximum_pool_size="4GiB"
... )
>>> rmm.mr.set_current_device_resource(pool)

Other MemoryResources include:

  • FixedSizeMemoryResource for allocating fixed blocks of memory
  • BinningMemoryResource for allocating blocks within specified "bin" sizes from different memory resources

MemoryResources are highly configurable and can be composed together in different ways. See help(rmm.mr) for more information.

Using RMM with third-party libraries

Using RMM with CuPy

You can configure CuPy to use RMM for memory allocations by setting the CuPy CUDA allocator to rmm_cupy_allocator:

>>> from rmm.allocators.cupy import rmm_cupy_allocator
>>> import cupy
>>> cupy.cuda.set_allocator(rmm_cupy_allocator)

Note: This only configures CuPy to use the current RMM resource for allocations. It does not initialize nor change the current resource, e.g., enabling a memory pool. See here for more information on changing the current memory resource.

Using RMM with Numba

You can configure Numba to use RMM for memory allocations using the Numba EMM Plugin.

This can be done in two ways:

  1. Setting the environment variable NUMBA_CUDA_MEMORY_MANAGER:
$ NUMBA_CUDA_MEMORY_MANAGER=rmm.allocators.numba python (args)
  1. Using the set_memory_manager() function provided by Numba:
>>> from numba import cuda
>>> from rmm.allocators.numba import RMMNumbaManager
>>> cuda.set_memory_manager(RMMNumbaManager)

Note: This only configures Numba to use the current RMM resource for allocations. It does not initialize nor change the current resource, e.g., enabling a memory pool. See here for more information on changing the current memory resource.

Using RMM with PyTorch

PyTorch can use RMM for memory allocation. For example, to configure PyTorch to use an RMM-managed pool:

import rmm
from rmm.allocators.torch import rmm_torch_allocator
import torch

rmm.reinitialize(pool_allocator=True)
torch.cuda.memory.change_current_allocator(rmm_torch_allocator)

PyTorch and RMM will now share the same memory pool.

You can, of course, use a custom memory resource with PyTorch as well:

import rmm
from rmm.allocators.torch import rmm_torch_allocator
import torch

# note that you can configure PyTorch to use RMM either before or
# after changing RMM's memory resource.  PyTorch will use whatever
# memory resource is configured to be the "current" memory resource at
# the time of allocation.
torch.cuda.change_current_allocator(rmm_torch_allocator)

# configure RMM to use a managed memory resource, wrapped with a
# statistics resource adaptor that can report information about the
# amount of memory allocated:
mr = rmm.mr.StatisticsResourceAdaptor(rmm.mr.ManagedMemoryResource())
rmm.mr.set_current_device_resource(mr)

x = torch.tensor([1, 2]).cuda()

# the memory resource reports information about PyTorch allocations:
mr.allocation_counts
Out[6]:
{'current_bytes': 16,
 'current_count': 1,
 'peak_bytes': 16,
 'peak_count': 1,
 'total_bytes': 16,
 'total_count': 1}

Taking ownership of C++ objects from Python.

When interacting with a C++ library that uses RMM from Python, one must be careful when taking ownership of rmm::device_buffer objects on the Python side. The rmm::device_buffer does not contain an owning reference to the memory resource used for its allocation (only a device_async_resource_ref), and the allocating user is expected to keep this memory resource alive for at least the lifetime of the buffer. When taking ownership of such a buffer in Python, we have no way (in the general case) of ensuring that the memory resource will outlive the buffer we are now holding.

To avoid any issues, we need two things:

  1. The C++ library we are interfacing with should accept a memory resource that is used for allocations that are returned to the user.
  2. When calling into the library from python, we should provide a memory resource whose lifetime we control. This memory resource should then be provided when we take ownership of any allocated rmm::device_buffers.

For example, suppose we have a C++ function that allocates device_buffers, which has a utility overload that defaults the memory resource to the current device resource:

std::unique_ptr<rmm::device_buffer> allocate(
  std::size_t size,
  rmm::mr::device_async_resource_ref mr = get_current_device_resource())
{
    return std::make_unique<rmm::device_buffer>(size, rmm::cuda_stream_default, mr);
}

The Python DeviceBuffer class has a convenience Cython function, c_from_unique_ptr to construct a DeviceBuffer from a unique_ptr<rmm::device_buffer>, taking ownership of it. To do this safely, we must ensure that the allocation that was done on the C++ side uses a memory resource we control. So:

# Bad, doesn't control lifetime
buffer_bad = DeviceBuffer.c_from_unique_ptr(allocate(10))

# Good, allocation happens with a memory resource we control
# mr is a DeviceMemoryResource
buffer_good = DeviceBuffer.c_from_unique_ptr(
    allocate(10, mr.get_mr()),
    mr=mr,
)

Note two differences between the bad and good cases:

  1. In the good case we pass the memory resource to the allocation function.
  2. In the good case, we pass the same memory resource to the DeviceBuffer constructor so that its lifetime is tied to the lifetime of the buffer.

Potential pitfalls of relying on get_current_device_resource

Functions in both the C++ and Python APIs that perform allocation typically default the memory resource argument to the value of get_current_device_resource. This is to simplify the interface for callers. When using a C++ library from Python, this defaulting is safe, as long as it is only the Python process that ever calls set_current_device_resource.

This is because the current device resource on the C++ side has a lifetime which is expected to be managed by the user. The resources set by rmm::mr::set_current_device_resource are stored in a static std::map whose keys are device ids and values are raw pointers to the memory resources. Consequently, rmm::mr::get_current_device_resource returns an object with no lifetime provenance. This is, for the reasons discussed above, not usable from Python. To handle this on the Python side, the Python-level set_current_device_resource sets the C++ resource and stores the Python object in a static global dictionary. The Python get_current_device_resource then does not use rmm::mr::get_current_device_resource and instead looks up the current device resource in this global dictionary.

Hence, if the C++ library we are interfacing with calls rmm::mr::set_current_device_resource, the C++ and Python sides of the program can disagree on what get_current_device_resource returns. The only safe thing to do if using the simplified interfaces is therefore to ensure that set_current_device_resource is only ever called on the Python side.