oneAPI Collective Communications Library (oneCCL)

Installation   |   Usage   |   Release Notes   |   Documentation   |   How to Contribute   |   License

oneAPI Collective Communications Library (oneCCL) provides an efficient implementation of communication patterns used in deep learning.

oneCCL is integrated into:

• Horovod* (distributed training framework). Refer to Horovod with oneCCL for details.
• PyTorch* (machine learning framework). Refer to PyTorch bindings for oneCCL for details.

oneCCL is governed by the UXL Foundation and is an implementation of the oneAPI specification.

Table of Contents

• Prerequisites
• Installation
• Usage
  • Launching Example Application
    • Using external MPI
  • Setting workers affinity
    • Automatic setup
    • Explicit setup
  • Using oneCCL package from CMake
    • oneCCLConfig files generation
• Governance
• Additional Resources
  • Blog Posts
  • Workshop Materials
• Contribute
• License
• Security Policy

Prerequisites

See System Requirements to learn about hardware and software requirements before getting started with oneCCL.

Installation

See install instructions to familiarize yourself with the general installation scenario and the customizations you can make during the build process.

Usage

Launch an Example Application

Use the command:

$ source <install_dir>/env/setvars.sh
$ mpirun -n 2 <install_dir>/examples/benchmark/benchmark

Use External mpi

In the vars.sh file, the ccl-bundled-mpi flag can have values "yes" or "no" to control whether bundled Intel MPI should be used or not. Current default value is "yes", which means that oneCCL temporarily overrides the mpi implementation in use.

In order to suppress the behavior and use user-supplied or system-default mpi, use the following command instead of sourcing setvars.sh:

$ source <install_dir>/env/vars.sh --ccl-bundled-mpi=no

The mpi implementation will not be overridden. In this case, you need to ensure that the system finds all required mpi-related binaries.

Set Workers Affinity

There are two ways to set worker threads (workers) affinity: automatically and explicitly.

Automatic Setup

1. Set the CCL_WORKER_COUNT environment variable with the desired number of workers per process.
2. Set the CCL_WORKER_AFFINITY environment variable with the value auto.

Example:

export CCL_WORKER_COUNT=4
export CCL_WORKER_AFFINITY=auto

With the variables above, oneCCL will create four workers per process and the pinning will depend from process launcher.

If an application is launched using mpirun that is provided by oneCCL distribution package, then workers will be automatically pinned to the last four cores available for the launched process. The exact IDs of CPU cores can be controlled by the mpirun parameters.

Otherwise, workers will be automatically pinned to the last four cores available on the node.

---

Explicit Setup

1. Set the CCL_WORKER_COUNT environment variable with the desired number of workers per process.
2. Set the CCL_WORKER_AFFINITY environment variable with the IDs of cores to pin local workers.

Example:

export CCL_WORKER_COUNT=4
export CCL_WORKER_AFFINITY=3,4,5,6

With the variables above, oneCCL will create four workers per process and pin them to the cores with the IDs of 3, 4, 5, and 6 respectively.

Use oneCCL package from CMake

oneCCLConfig.cmake and oneCCLConfigVersion.cmake are included into oneCCL distribution.

With these files, you can integrate oneCCL into a project with the find_package command. Successful invocation of find_package(oneCCL <options>) creates imported target oneCCL that can be passed to the target_link_libraries command.

For example:

project(Foo)
add_executable(foo foo.cpp)

# Search for oneCCL
find_package(oneCCL REQUIRED)

# Connect oneCCL to foo
target_link_libraries(foo oneCCL)

oneCCLConfig Files Generation

To generate oneCCLConfig files for oneCCL package, use the provided cmake/scripts/config_generation.cmake file:

cmake [-DOUTPUT_DIR=<output_dir>] -P cmake/script/config_generation.cmake

OS File Descriptors

oneCCL uses Level Zero IPC handles so that a process can access a memory allocation done by a different process. However, these IPC handles consume OS File Descriptors (FDs). To avoid running out of OS FDs, we recommend to increase the default limit of FDs in the system for applications running with oneCCL and GPU buffers.

The number of FDs required is application-dependent, but the recommended limit is 1048575. This value can be modified with the ulimit command.

Governance

The oneCCL project is governed by the UXL Foundation and you can get involved in this project in multiple ways. It is possible to join the Special Interest Groups (SIG) meetings where the group discusses and demonstrates work using the foundation projects. Members can also join the Open Source and Specification Working Group meetings.

You can also join the mailing lists for the UXL Foundation to be informed of when meetings are happening and receive the latest information and discussions.

Additional Resources

Blog Posts

• Optimizing DLRM by using PyTorch with oneCCL Backend
• Intel MLSL Makes Distributed Training with MXNet Faster

Workshop Materials

• oneAPI, oneCCL and OFI: Path to Heterogeneous Architecure Programming with Scalable Collective Communications: recording and slides

Contribute

See CONTRIBUTING for more information.

License

Distributed under the Apache License 2.0 license. See LICENSE for more information.

Security Policy

See SECURITY for more information.