KnetML / singularity-images

This repo contains a simgularity recipe for setting-up and running Knet easily.

Singularity Image for Knet

USAGE

• Pull the base image

singularity pull --name knet.simg shub://KnetML/singularity-images:latest

• Not to loose your changes, you should create an overlay image.

singularity image.create --size 2048 overlay.simg

• If you need more space, you can expand this image later

singularity image.expand overlay.simg

• Connect to the container with the overlay image.

singularity shell --overlay overlay.simg Knet.simg

• If you want to enable the gpu usage, you can bind your cuda and cudnn paths.

singularity shell --bind <your cuda path>:/usr/local/cuda,<your cudnn path>:/opt/cudnn --overlay overlay.simg --nv Knet.simg

• Build Knet

julia -e 'Pkg.build("Knet")'

• Now you can work under the /workdir folder

Running the image

You can also run the image directly. It starts an ipython notebook server. You can connect to this server on localhost:8888.

singularity run --bind <your cuda path>:/usr/local/cuda,<your cudnn path>:/opt/cudnn --overlay overlay.simg --nv Knet.simg

Usage on the kuacc cluster

• Load necessary modules

module load cuda/9.0
module load cudnn/7.0.4/cuda-9.0
module load singularity

• Set SINGULARITY_CACHEDIR as your scratch

export SINGULARITY_CACHEDIR=/scratch/users/username

• On a compute node with gpu

singularity shell --bind /usr/local/cuda-9.0/:/usr/local/cuda,/kuacc/apps/cudnn/v7.0.4_CUDA_9.0:/opt/cudnn --overlay overlay.img --nv Knet.simg

DEPENDENCIES

• Singularity
• Cuda drivers and libraries (optional)
• Cudnn library (optinal)