RCCL

ROCm Communication Collectives Library

Note: The published documentation is available at RCCL in an organized easy-to-read format that includes a table of contents and search functionality. The documentation source files reside in the rccl/docs folder in this repository. As with all ROCm projects, the documentation is open source. For more information, see Contribute to ROCm documentation.

Introduction

RCCL (pronounced "Rickle") is a stand-alone library of standard collective communication routines for GPUs, implementing all-reduce, all-gather, reduce, broadcast, reduce-scatter, gather, scatter, and all-to-all. There is also initial support for direct GPU-to-GPU send and receive operations. It has been optimized to achieve high bandwidth on platforms using PCIe, xGMI as well as networking using InfiniBand Verbs or TCP/IP sockets. RCCL supports an arbitrary number of GPUs installed in a single node or multiple nodes, and can be used in either single- or multi-process (e.g., MPI) applications.

The collective operations are implemented using ring and tree algorithms and have been optimized for throughput and latency. For best performance, small operations can be either batched into larger operations or aggregated through the API.

Requirements

1. ROCm supported GPUs
2. ROCm stack installed on the system (HIP runtime & HIP-Clang)

Quickstart RCCL Build

RCCL directly depends on HIP runtime plus the HIP-Clang compiler, which are part of the ROCm software stack. For ROCm installation instructions, see https://github.com/ROCm/ROCm.

The root of this repository has a helper script install.sh to build and install RCCL with a single command. It hard-codes configurations that can be specified through invoking cmake directly, but it's a great way to get started quickly and can serve as an example of how to build/install RCCL.

To build the library using the install script:

./install.sh

For more info on build options/flags when using the install script, use ./install.sh --help

./install.sh --help
RCCL build & installation helper script
 Options:
       --address-sanitizer     Build with address sanitizer enabled
    -d|--dependencies          Install RCCL depdencencies
       --debug                 Build debug library
       --enable_backtrace      Build with custom backtrace support
       --disable-colltrace     Build without collective trace
       --disable-msccl-kernel  Build without MSCCL kernels
       --disable-mscclpp       Build without MSCCL++ support
    -f|--fast                  Quick-build RCCL (local gpu arch only, no backtrace, and collective trace support)
    -h|--help                  Prints this help message
    -i|--install               Install RCCL library (see --prefix argument below)
    -j|--jobs                  Specify how many parallel compilation jobs to run ($nproc by default)
    -l|--local_gpu_only        Only compile for local GPU architecture
       --amdgpu_targets        Only compile for specified GPU architecture(s). For multiple targets, seperate by ';' (builds for all supported GPU architectures by default)
       --no_clean              Don't delete files if they already exist
       --npkit-enable          Compile with npkit enabled
       --openmp-test-enable    Enable OpenMP in rccl unit tests
       --roctx-enable          Compile with roctx enabled (example usage: rocprof --roctx-trace ./rccl-program)
    -p|--package_build         Build RCCL package
       --prefix                Specify custom directory to install RCCL to (default: `/opt/rocm`)
       --rm-legacy-include-dir Remove legacy include dir Packaging added for file/folder reorg backward compatibility
       --run_tests_all         Run all rccl unit tests (must be built already)
    -r|--run_tests_quick       Run small subset of rccl unit tests (must be built already)
       --static                Build RCCL as a static library instead of shared library
    -t|--tests_build           Build rccl unit tests, but do not run
       --time-trace            Plot the build time of RCCL (requires `ninja-build` package installed on the system)
       --verbose               Show compile commands

By default, RCCL builds for all GPU targets defined in DEFAULT_GPUS in CMakeLists.txt. To target specific GPU(s), and potentially reduce build time, use --amdgpu_targets as a ; separated string listing GPU(s) to target.

Manual build

To build the library using CMake:

$ git clone https://github.com/ROCm/rccl.git --recurse-submodules
$ cd rccl
$ mkdir build
$ cd build
$ cmake ..
$ make -j 16      # Or some other suitable number of parallel jobs

If you have already cloned, you can checkout the mscclpp submodule manually.

$ cd ext-src/mscclpp
$ git submodule update --init --recursive

You may substitute an installation path of your own choosing by passing CMAKE_INSTALL_PREFIX. For example:

$ cmake -DCMAKE_INSTALL_PREFIX=$PWD/rccl-install ..

Note: ensure rocm-cmake is installed, apt install rocm-cmake.

To build the RCCL package and install package :

Assuming you have already cloned this repository and built the library as shown in the previous section:

$ cd rccl/build
$ make package
$ sudo dpkg -i *.deb

RCCL package install requires sudo/root access because it creates a directory called "rccl" under /opt/rocm/. This is an optional step and RCCL can be used directly by including the path containing librccl.so.

Enabling peer-to-peer transport

In order to enable peer-to-peer access on machines with PCIe-connected GPUs, the HSA environment variable HSA_FORCE_FINE_GRAIN_PCIE=1 is required to be set, on top of requiring GPUs that support peer-to-peer access and proper large BAR addressing support.

Tests

There are rccl unit tests implemented with the Googletest framework in RCCL. The rccl unit tests require Googletest 1.10 or higher to build and execute properly (installed with the -d option to install.sh). To invoke the rccl unit tests, go to the build folder, then the test subfolder, and execute the appropriate rccl unit test executable(s).

rccl unit test names are now of the format:

CollectiveCall.[Type of test]

Filtering of rccl unit tests should be done with environment variable and by passing the --gtest_filter command line flag, for example:

UT_DATATYPES=ncclBfloat16 UT_REDOPS=prod ./rccl-UnitTests --gtest_filter="AllReduce.C*"

will run only AllReduce correctness tests with float16 datatype. A list of available filtering environment variables appears at the top of every run. See "Running a Subset of the Tests" at https://google.github.io/googletest/advanced.html#running-a-subset-of-the-tests for more information on how to form more advanced filters.

There are also other performance and error-checking tests for RCCL. These are maintained separately at https://github.com/ROCm/rccl-tests. See the rccl-tests README for more information on how to build and run those tests.

NPKit

RCCL integrates NPKit, a profiler framework that enables collecting fine-grained trace events in RCCL components, especially in giant collective GPU kernels.

Please check NPKit sample workflow for RCCL as a fully automated usage example. It also provides good templates for the following manual instructions.

To manually build RCCL with NPKit enabled, pass -DNPKIT_FLAGS="-DENABLE_NPKIT -DENABLE_NPKIT_...(other NPKit compile-time switches)" with cmake command. All NPKit compile-time switches are declared in the RCCL code base as macros with prefix ENABLE_NPKIT_, and they control which information will be collected. Also note that currently NPKit only supports collecting non-overlapped events on GPU, and -DNPKIT_FLAGS should follow this rule.

To manually run RCCL with NPKit enabled, environment variable NPKIT_DUMP_DIR needs to be set as the NPKit event dump directory. Also note that currently NPKit only supports 1 GPU per process.

To manually analyze NPKit dump results, please leverage npkit_trace_generator.py.

MSCCL/MSCCL++

RCCL integrates MSCCL and MSCCL++ to leverage the highly efficient GPU-GPU communication primitives for collective operations. Thanks to Microsoft Corporation for collaborating with us in this project.

MSCCL uses XMLs for different collective algorithms on different architectures. RCCL collectives can leverage those algorithms once the corresponding XML has been provided by the user. The XML files contain the sequence of send-recv and reduction operations to be executed by the kernel. On MI300X, MSCCL is enabled by default. On other platforms, the users may have to enable this by setting RCCL_MSCCL_FORCE_ENABLE=1. By default, MSCCL will only be used if every rank belongs to a unique process; to disable this restriction for multi-threaded or single-threaded configurations, set RCCL_MSCCL_ENABLE_SINGLE_PROCESS=1.

On the other hand, RCCL allreduce and allgather collectives can leverage the efficient MSCCL++ communication kernels for certain message sizes. MSCCL++ support is available whenever MSCCL support is available. Users need to set the RCCL environment variable RCCL_MSCCLPP_ENABLE=1 to run RCCL workload with MSCCL++ support. It is also possible to set the message size threshold for using MSCCL++ by using the environment variable RCCL_MSCCLPP_THRESHOLD. Once RCCL_MSCCLPP_THRESHOLD (the default value is 1MB) is set, RCCL will invoke MSCCL++ kernels for all message sizes less than or equal to the specified threshold.

If some restrictions are not met, it will fall back to MSCCL or RCCL. The following are restrictions on using MSCCL++:

• Message size must be a non-zero multiple of 32 bytes
• Does not support hipMallocManaged buffers
• Allreduce only supports float16, int32, uint32, float32, and bfloat16 data types
• Allreduce only supports the sum op

Library and API Documentation

Please refer to the RCCL Documentation Site for current documentation.

How to build documentation

Run the steps below to build documentation locally.

cd docs
pip3 install -r sphinx/requirements.txt
python3 -m sphinx -T -E -b html -d _build/doctrees -D language=en . _build/html

Improving performance on MI300 when using less than 8 GPUs

On a system with 8*MI300X GPUs, each pair of GPUs are connected with dedicated XGMI links in a fully-connected topology. So, for collective operations, one can achieve good performance when all 8 GPUs (and all XGMI links) are used. When using less than 8 GPUs, one can only achieve a fraction of the potential bandwidth on the system.

But, if your workload warrants using less than 8 MI300 GPUs on a system, you can set the run-time variable NCCL_MIN_NCHANNELS to increase the number of channels.
E.g.: export NCCL_MIN_NCHANNELS=32

Increasing the number of channels can be beneficial to performance, but it also increases GPU utilization for collective operations.

Additionally, we have pre-defined higher number of channels when using only 2 GPUs or 4 GPUs on a 8*MI300 system. Here, RCCL will use 32 channels for the 2 MI300 GPUs scenario and 24 channels for the 4 MI300 GPUs scenario.

Copyright

All source code and accompanying documentation is copyright (c) 2015-2022, NVIDIA CORPORATION. All rights reserved.

All modifications are copyright (c) 2019-2022 Advanced Micro Devices, Inc. All rights reserved.