The Unified Memory Framework (UMF) is a library for constructing allocators and memory pools. It also contains broadly useful abstractions and utilities for memory management. UMF allows users to manage multiple memory pools characterized by different attributes, allowing certain allocation types to be isolated from others and allocated using different hardware resources as required.

For a quick introduction to UMF usage, please see examples documentation, which includes the code of the basic example. There are also more advanced examples that allocate USM memory from the Level Zero device using the Level Zero API and UMF Level Zero memory provider and CUDA device using the CUDA API and UMF CUDA memory provider.

UMF's experimental CTL API is showcased in the CTL example, which explores provider and pool statistics, and in the custom CTL example, which wires CTL support into a custom memory provider. These examples rely on experimental headers which may change in future releases.

Required packages:

• libhwloc-dev >= 2.3.0 (Linux) / hwloc >= 2.3.0 (Windows)
• C compiler
• CMake >= 3.14.0

For development and contributions:

• clang-format-15.0 (can be installed with python -m pip install clang-format==15.0.7)
• cmake-format-0.6 (can be installed with python -m pip install cmake-format==0.6.13)
• black (can be installed with python -m pip install black==24.3.0)

Note: All devs dependencies are defined in third-party/requirements.txt and can be installed, for example: pip install -r third_party/requirements.txt.

For building tests and multithreaded benchmarks:

• C++ compiler with C++17 support

For Level Zero memory provider tests:

• Level Zero headers and libraries
• compatible GPU with installed driver

Executable and binaries will be in build/bin. The {build_config} can be either Debug or Release.

cmake -B build -DCMAKE_BUILD_TYPE={build_config}
cmake --build build -j $(nproc)

Generating Visual Studio Project. EXE and binaries will be in build/bin/{build_config}. The {build_config} can be either Debug or Release.

cmake -B build -G "Visual Studio 15 2017 Win64"
cmake --build build --config {build_config} -j $Env:NUMBER_OF_PROCESSORS

UMF comes with a single-threaded micro benchmark based on ubench. In order to build the benchmark, the UMF_BUILD_BENCHMARKS CMake configuration flag has to be turned ON.

UMF also provides multithreaded benchmarks that can be enabled by setting both UMF_BUILD_BENCHMARKS and UMF_BUILD_BENCHMARKS_MT CMake configuration flags to ON. Multithreaded benchmarks require C++ support.

The Scalable Pool requirements can be found in the relevant 'Memory Pool managers' section below.

List of sanitizers available on Linux:

• AddressSanitizer
• UndefinedBehaviorSanitizer
• ThreadSanitizer
  • Is mutually exclusive with other sanitizers.
• MemorySanitizer
  • Requires linking against MSan-instrumented libraries to prevent false positive reports. More information here.

List of sanitizers available on Windows:

• AddressSanitizer

Listed sanitizers can be enabled with appropriate CMake options.

To enable fuzz testing, the UMF_BUILD_FUZZTESTS CMake configuration flag must be set to ON. Note, that this feature is supported only on Linux and requires Clang. Additionally, ensure that the CMAKE_PREFIX_PATH includes the directory containing the libraries necessary for fuzzing (e.g., Clang's libclang_rt.fuzzer_no_main-x86_64.a).

Example:

cmake -B build -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_BUILD_TYPE=Debug -DUMF_BUILD_FUZZTESTS=ON -DCMAKE_PREFIX_PATH=/path/to/fuzzer/libs

List of options provided by CMake:

A UMF memory pool is a combination of a pool allocator and a memory provider. A memory provider is responsible for coarse-grained memory allocations and management of memory pages, while the pool allocator controls memory pooling and handles fine-grained memory allocations.

Pool allocator can leverage existing allocators (e.g. jemalloc or tbbmalloc) or be written from scratch.

UMF comes with predefined pool allocators (see include/umf/pools) and providers (see include/umf/providers). UMF can also work with user-defined pools and providers that implement a specific interface (see include/umf/memory_pool_ops.h and include/umf/memory_provider_ops.h).

More detailed documentation is available here: https://oneapi-src.github.io/unified-memory-framework/

A memory provider that can provide memory from a given pre-allocated buffer.

A memory provider that provides memory from an operating system.

OS memory provider supports two types of memory mappings (set by the visibility parameter):

1. private memory mapping (UMF_MEM_MAP_PRIVATE)
2. shared memory mapping (UMF_MEM_MAP_SHARED - supported on Linux only yet)

IPC API requires the UMF_MEM_MAP_SHARED memory visibility mode (UMF_RESULT_ERROR_INVALID_ARGUMENT is returned otherwise).

IPC API uses file descriptor duplication, which requires the pidfd_getfd(2) system call to obtain a duplicate of another process's file descriptor. This system call is supported since Linux 5.6. Required permission ("restricted ptrace") is governed by the PTRACE_MODE_ATTACH_REALCREDS check (see ptrace(2)). To allow file descriptor duplication in a binary that opens IPC handle, you can call prctl(PR_SET_PTRACER, ...) in the producer binary that gets the IPC handle. Alternatively you can change the ptrace_scope globally in the system, e.g.:

sudo bash -c "echo 0 > /proc/sys/kernel/yama/ptrace_scope"

There are available two mechanisms for the shared memory mapping:

1. a named shared memory object (used if the shm_name parameter is not NULL) or
2. an anonymous file descriptor (used if the shm_name parameter is NULL)

The shm_name parameter should be a null-terminated string of up to NAME_MAX (i.e., 255) characters none of which are slashes.

An anonymous file descriptor for the shared memory mapping will be created using:

1. memfd_secret() syscall - (if it is implemented and) if the UMF_MEM_FD_FUNC environment variable does not contain the "memfd_create" string or
2. memfd_create() syscall - otherwise (and if it is implemented).

IPC API on Linux requires the PTRACE_MODE_ATTACH_REALCREDS permission (see ptrace(2)) to duplicate another process's file descriptor (see above).

Packages required for tests (Linux-only yet):

• libnuma-dev

A memory provider that provides memory from L0 device.

IPC API uses file descriptor duplication, which requires the pidfd_getfd(2) system call to obtain a duplicate of another process's file descriptor. This system call is supported since Linux 5.6. Required permission ("restricted ptrace") is governed by the PTRACE_MODE_ATTACH_REALCREDS check (see ptrace(2)). To allow file descriptor duplication in a binary that opens IPC handle, you can call prctl(PR_SET_PTRACER, ...) in the producer binary that gets the IPC handle. Alternatively you can change the ptrace_scope globally in the system, e.g.:

sudo bash -c "echo 0 > /proc/sys/kernel/yama/ptrace_scope"

1. Linux or Windows OS
2. The UMF_BUILD_LEVEL_ZERO_PROVIDER option turned ON (by default)
3. IPC API on Linux requires the PTRACE_MODE_ATTACH_REALCREDS permission (see ptrace(2)) to duplicate another process's file descriptor (see above).

Additionally, required for tests:

4. The UMF_BUILD_GPU_TESTS option turned ON
5. System with Level Zero compatible GPU
6. Required packages:
   • liblevel-zero-dev (Linux) or level-zero-sdk (Windows)

A memory provider that provides memory from a device DAX (a character device file like /dev/daxX.Y). It can be used when large memory mappings are needed.

1. Linux OS
2. A character device file /dev/daxX.Y created in the OS.

A memory provider that provides memory by mapping a regular, extendable file.

IPC API requires the UMF_MEM_MAP_SHARED memory visibility mode (UMF_RESULT_ERROR_INVALID_ARGUMENT is returned otherwise).

The memory visibility mode parameter must be set to UMF_MEM_MAP_SHARED in case of FSDAX.

1. Linux OS
2. A length of a path of a file to be mapped can be PATH_MAX (4096) characters at most.

A memory provider that provides memory from CUDA device.

1. Linux or Windows OS
2. The UMF_BUILD_CUDA_PROVIDER option turned ON (by default)

Additionally, required for tests:

3. The UMF_BUILD_GPU_TESTS option turned ON
4. System with CUDA compatible GPU
5. Required packages:
   • nvidia-cuda-dev (Linux) or cuda-sdk (Windows)

This memory pool is distributed as part of libumf. It forwards all requests to the underlying memory provider. Currently umfPoolRealloc, umfPoolCalloc and umfPoolMallocUsableSize functions are not supported by the proxy pool.

The Disjoint pool is designed to keep internal metadata separate from user data. This separation is particularly useful when user data needs to be placed in memory with relatively high latency, such as GPU memory or disk storage.

Jemalloc pool is a jemalloc-based memory pool manager built as a separate static library: libjemalloc_pool.a on Linux and jemalloc_pool.lib on Windows. The UMF_BUILD_LIBUMF_POOL_JEMALLOC option has to be turned ON to build this library.

jemalloc is required to build the jemalloc pool.

In case of Linux OS jemalloc is built from the (fetched) sources with the following non-default options enabled:

• --with-jemalloc-prefix=je_ - adds the je_ prefix to all public APIs,
• --disable-cxx - disables C++ integration, it will cause the new and the delete operators implementations to be omitted.
• --disable-initial-exec-tls - disables the initial-exec TLS model for jemalloc's internal thread-local storage (on those platforms that support explicit settings), it can allow jemalloc to be dynamically loaded after program startup (e.g. using dlopen()).

The default jemalloc package is required on Windows.

1. The UMF_BUILD_LIBUMF_POOL_JEMALLOC option turned ON
2. jemalloc is required:

• on Linux and MacOS: jemalloc is fetched and built from sources (a custom build),
• on Windows: the default jemalloc package is required

Scalable Pool is a oneTBB-based memory pool manager. It is distributed as part of libumf. To use this pool, TBB must be installed in the system.

Packages required for using this pool and executing tests/benchmarks (not required for build):

• libtbb-dev (libtbbmalloc.so.2) on Linux or tbb (tbbmalloc.dll) on Windows

Note: The memspace, memtarget and mempolicy APIs are experimental and may change in future releases.

TODO: Add general information about memspaces.

Memspace backed by all available NUMA nodes discovered on the platform. Can be retrieved using umfMemspaceHostAllGet.

Memspace backed by all available NUMA nodes discovered on the platform sorted by capacity. Can be retrieved using umfMemspaceHighestCapacityGet.

Memspace backed by an aggregated list of NUMA nodes identified as highest bandwidth after selecting each available NUMA node as the initiator. Querying the bandwidth value requires HMAT support on the platform. Calling umfMemspaceHighestBandwidthGet() will return NULL if it's not supported.

Memspace backed by an aggregated list of NUMA nodes identified as lowest latency after selecting each available NUMA node as the initiator. Querying the latency value requires HMAT support on the platform. Calling umfMemspaceLowestLatencyGet() will return NULL if it's not supported.

UMF provides the UMF proxy library (umf_proxy) that makes it possible to override the default allocator in other programs in both Linux and Windows.

To enable this feature, the UMF_BUILD_SHARED_LIBRARY option needs to be turned ON.

In case of Linux it can be done without any code changes using the LD_PRELOAD environment variable:

LD_PRELOAD=/usr/lib/libumf_proxy.so myprogram

The memory used by the proxy memory allocator is mmap'ed:

1. with the MAP_PRIVATE flag by default or
2. with the MAP_SHARED flag if the UMF_PROXY environment variable contains one of two following strings: page.disposition=shared-shm or page.disposition=shared-fd. These two options differ in a mechanism used during IPC:
   • page.disposition=shared-shm - IPC uses the named shared memory. An SHM name is generated using the umf_proxy_lib_shm_pid_$PID pattern, where $PID is the PID of the process. It creates the /dev/shm/umf_proxy_lib_shm_pid_$PID file.
   • page.disposition=shared-fd - IPC API uses file descriptor duplication, which requires the pidfd_getfd(2) system call to obtain a duplicate of another process's file descriptor. This system call is supported since Linux 5.6. Required permission ("restricted ptrace") is governed by the PTRACE_MODE_ATTACH_REALCREDS check (see ptrace(2)). To allow file descriptor duplication in a binary that opens IPC handle, you can call prctl(PR_SET_PTRACER, ...) in the producer binary that gets the IPC handle. Alternatively you can change the ptrace_scope globally in the system, e.g.: sudo bash -c "echo 0 > /proc/sys/kernel/yama/ptrace_scope".

Size threshold

The size threshold feature (Linux only) causes that all allocations of size less than the given threshold value go to the default system allocator instead of the proxy library. It can be enabled by adding the size.threshold=<value> string to the UMF_PROXY environment variable (with ';' as a separator), for example: UMF_PROXY="page.disposition=shared-shm;size.threshold=64".

Remark: changing a size of allocation (using realloc() ) does not change the allocator (realloc(malloc(threshold - 1), threshold + 1) still belongs to the default system allocator and realloc(malloc(threshold + 1), threshold - 1) still belongs to the proxy library pool allocator).

In case of Windows it requires:

1. explicitly linking your program dynamically with the umf_proxy.dll library
2. (C++ code only) including proxy_lib_new_delete.h in a single(!) source file in your project to override also the new/delete operations.

All contributions to the UMF project are most welcome! Before submitting an issue or a Pull Request, please read Contribution Guide.

To enable logging in UMF source files please follow the guide in the web documentation.

Integration of UMF into another project via CMake's FetchContent, is possible with:

include(FetchContent)

FetchContent_Declare(
    unified-memory-framework
    GIT_REPOSITORY https://github.com/oneapi-src/unified-memory-framework.git
    GIT_TAG main  # This will pull the latest (potentially unstable) changes from the main branch
)
FetchContent_MakeAvailable(unified-memory-framework)

add_executable(some_example some_example.cpp)
target_include_directories(some_example PRIVATE ${unified-memory-framework_SOURCE_DIR}/include)
target_link_libraries(some_example PRIVATE umf::umf umf::headers)

The contents of this repository may have been developed with support from one or more Intel-operated generative artificial intelligence solutions.