Installation Guide

Requirements

ChainerMN depends on the following software libraries: CUDA-Aware MPI, NVIDIA NCCL, and a few Python packages including CuPy and MPI4py.

Note

In Chainer v5, ChainerMN became a part of Chainer package. Installing Chainer (pip install chainer) automatically makes ChainerMN available.

Before upgrading from Chainer v4 to v5 or later, make sure to remove existing chainermn package (pip uninstall chainermn).

CUDA-Aware MPI

ChainerMN relies on MPI. In particular, for efficient communication between GPUs, it uses CUDA-aware MPI. For details about CUDA-aware MPI, see this introduction article. (If you use only the CPU mode, MPI does not need to be CUDA-Aware. See Installation on Non-GPU Environments for more details.)

The CUDA-aware features depend on several MPI packages, which need to be configured and built properly. The following are examples of Open MPI and MVAPICH.

Open MPI (for details, see Open MPI’s official instructions):

$ ./configure --with-cuda
$ make -j4
$ sudo make install

MVAPICH (for details, see Mvapich’s official instructions):

$ ./configure --enable-cuda
$ make -j4
$ sudo make install
$ export MV2_USE_CUDA=1  # Should be set all the time when using ChainerMN

NCCL

Note

If you are installing CuPy using wheels (i.e., pip install cupy-cudaXX where XX is the CUDA version), you don’t have to install NCCL manually. The latest NCCL 2.x library is bundled with CuPy wheels.

See CuPy Installation Guide for the detailed steps to install CuPy.

To enable efficient intra- and inter-node GPU-to-GPU communication, we use NVIDIA Collective Communications Library (NCCL). See NCCL’s official instructions for installation.

ChainerMN requires NCCL even if you have only one GPU per node. The only exception is when you run ChainerMN on CPU-only environments. See Installation on Non-GPU Environments for more details.

Note

We reccomend NCCL 2 but NCCL 1 can be used. When you use CUDA 7.0 and 7.5, please install NCCL 1 because NCCL 2 is not supported with CUDA 7.0 and 7.5. However, for NCCL 1, PureNcclCommunicator is not supported in ChainerMN. If you use NCCL 1, please properly configure environment variables to expose NCCL both when you install and use ChainerMN. Typical configurations should look like the following:

export NCCL_ROOT=<path to NCCL directory>
export CPATH=$NCCL_ROOT/include:$CPATH
export LD_LIBRARY_PATH=$NCCL_ROOT/lib/:$LD_LIBRARY_PATH
export LIBRARY_PATH=$NCCL_ROOT/lib/:$LIBRARY_PATH

If you change the version of NCCL installed, you have to reinstall CuPy. Because, current ChainerMN applies CuPy to use NCCL. See CuPy official instructions for reinstalltion.

MPI4py

ChainerMN depends on a few Python packages, which are automatically installed when you install ChainerMN.

However, among them, we need to be a little careful about MPI4py. It links to MPI at installation time, so please be sure to properly configure environment variables so that MPI is available at installation time. In particular, if you have multiple MPI implementations in your environment, please expose the implementation that you want to use both when you install and use ChainerMN.

CuPy

Chainer and ChainerMN rely on CuPy to use GPUs. Please refer to CuPy Installation Guide for the detailed steps to install CuPy. CuPy requires NCCL to be enabled. See Check if NCCL is enabled in CuPy, if you want to check whether NCCL is enabled in CuPy.

Chainer and ChainerMN can be installed without CuPy, in which case the corresponding features are not available. See Installation on Non-GPU Environments for more details.

Tested Environments

We tested ChainerMN on all the following environments.

  • OS
    • Ubuntu 14.04 LTS 64bit
    • Ubuntu 16.04 LTS 64bit
  • Python 2.7.13 3.5.1 3.6.1
  • MPI
    • openmpi 1.10.7 2.1.2
  • MPI4py 3.0.0
  • NCCL 2.2.13

Installation on Non-GPU Environments

Users who want to try ChainerMN in CPU-only environment may skip installation of CuPy. Non-GPU set up may not be performant as GPU-enabled set up, but would be useful for testing or debugging training program in non-GPU environment such as laptops or CI jobs.

In this case, the MPI does not have to be CUDA-aware. Only naive communicator works with the CPU mode.