Running NVCaffe

Before you can run an NGC deep learning framework container, your Docker environment must support NVIDIA GPUs. To run a container, issue the appropriate command as explained in the Running A Container chapter in the NVIDIA Containers And Frameworks User Guide and specify the registry, repository, and tags.

On a system with GPU support for NGC containers, the following occurs when running a container:
  • The Docker engine loads the image into a container which runs the software.
  • You define the runtime resources of the container by including additional flags and settings that are used with the command. These flags and settings are described in Running A Container.
  • The GPUs are explicitly defined for the Docker container (defaults to all GPUs, can be specified using NV_GPU environment variable).

The method implemented in your system depends on the DGX OS version installed (for DGX systems), the specific NGC Cloud Image provided by a Cloud Service Provider, or the software that you have installed in preparation for running NGC containers on TITAN PCs, Quadro PCs, or vGPUs.

  1. Issue the command for the applicable release of the container that you want. The following command assumes you want to pull the latest container.
    docker pull nvcr.io/nvidia/caffe:19.11-py3
  2. Open a command prompt and paste the pull command. The pulling of the container image begins. Ensure the pull completes successfully before proceeding to the next step.
  3. Run the container image. To run the container, choose interactive mode or non-interactive mode.
    1. Interactive mode: Open a command prompt and issue:
      docker run --gpus all -it --rm -v local_dir:container_dir 
      nvcr.io/nvidia/caffe:<xx.xx>-py3
    2. Non-interactive mode: Open a command prompt and issue:
      docker run --gpus all --rm -v local_dir:container_dir 
      nvcr.io/nvidia/caffe:<xx.xx>-py3 caffe train …
      

    You might want to pull in data and model descriptions from locations outside the container for use by NVCaffe or save results to locations outside the container. To accomplish this, the easiest method is to mount one or more host directories as Docker® data volumes.

    Note: In order to share data between ranks, NVIDIA® Collective Communications Library ™ (NCCL) may require shared system memory for IPC and pinned (page-locked) system memory resources. The operating system’s limits on these resources may need to be increased accordingly. Refer to your system’s documentation for details.
    In particular, Docker containers default to limited shared and pinned memory resources. When using NCCL inside a container, it is recommended that you increase these resources by issuing:
    --shm-size=1g --ulimit memlock=-1
    in the command line to:
    docker run --gpus all

Running An NVCaffe Container On A Cluster

NVCaffe supports training on multiple nodes using OpenMPI version 2.0 protocol, however, you cannot specify the number of threads per process because NVCaffe has its own thread manager (currently it runs one worker thread per GPU). For example:
dgx job submit --name jobname --volume <src>:<dst> --tasks 48 
--clusterid <id> --gpu 8 --cpu 64 --mem 480 --image <tag> --nc "mpirun 
-bind-to none -np 48 -pernode --tag-output caffe train --solver 
solver.prototxt --gpu all >> /logs/caffe.log 2>&1"