Tips for installing CUDA into a rocker docker container

seabbs · January 23, 2019, 1:31pm

Hi there,

This seems like something that people must be doing but I have been struggling with it. The use case is to make use of my GPU for some P-MCMC and machine learning (Xgboost).

So far I have got to the stage of having nvidia-docker (https://github.com/NVIDIA/nvidia-docker) working on my system (the example nvidia-smi container works fine).

Most of the suggested solutions I have seen recommended basing the container image on one of the nvidia supplied cuda images. The others recommend a manual install but I have been struggling to get this working. I would much prefer sticking with the rocker base image so any ideas would be great.

Thanks,

Sam

noamross · January 23, 2019, 1:51pm

I am way behind on a project to have a GPU-enabled rocker container. There are two approaches: First, build rocker on top of the NVIDIA images, which is easier for those who know the Rocker stack well. Example:

github.com

ecohealthalliance/reservoir/blob/52ff94b8a82fa854699edb44e1e4a23481f757b9/Dockerfile.gpu

FROM rocker/geospatial:latest
MAINTAINER "Noam Ross" ross@ecohealthalliance.org

## Install stuff
### Shiny
#RUN export ADD=shiny && bash /etc/cont-init.d/add
ADD latest-rstudio-preview.R /latest-rstudio-preview.R
## Installs to help install
RUN apt-get update && apt-get install -y --force-yes --allow-unauthenticated --no-install-recommends --no-upgrade \
   curl gnupg gnupg2 apt-transport-https dirmngr libclang-dev software-properties-common \
 && AZ_REPO=$(lsb_release -cs) \
 && echo "deb [arch=amd64] https://packages.microsoft.com/repos/azure-cli/ $AZ_REPO main" | \
     tee /etc/apt/sources.list.d/azure-cli.list \
 && curl -L https://packages.microsoft.com/keys/microsoft.asc | apt-key add - 
 
 
## Get docker itself
RUN curl -sSL https://get.docker.com/ | sh \
### Shell tools
 && apt-get update && apt-get install -y --force-yes --allow-unauthenticated --no-install-recommends --no-upgrade \

This file has been truncated. show original

Alternatively, install NVIDIA software into the Docker image. This beast of a Dockerfile installs the NVIDIA packages and GPU-enabled versions of tensorflow and xgboost (along with a ton of other things - it’s our group’s common working image, adopted for GPU). I’ve neglected it for a while and I’m not 100% it works in its current condition, but it was working at some point : https://github.com/ecohealthalliance/reservoir/blob/master/Dockerfile.gpu

noamross · January 24, 2019, 11:29pm

Whoops, I just realized I copied the same link twice. This is the one that builds rocker on top of the NVIDIA base: https://github.com/rocker-org/ml/blob/0f128aa97e12868d0a6c92bffc7a832aecce84d2/gpu/Dockerfile

seabbs · January 25, 2019, 5:38pm

Awesome thanks for this Noam,

Not knowing the rocker stack that well I went with the second approach.

Did some minor adjustments to the Dockerfile you linked too (which is all in working order by the way) but rebased to rocker/tidyverse.

If it helps anyone my modified version is here (currently not working with h2o): https://github.com/seabbs/tidyverse-gpu

Very much looking forward to an out the box GPU enabled rocker container! Just got my hands on a GPU - hopefully, it is worth the setup effort.

Thanks,

Sam

noamross · January 25, 2019, 5:57pm

Glad it helped and good to know it’s functioning!

Topic		Replies	Views
Using the GPU backend in h2o.xgboost in a rocker based Docker container General Q&A r	2	1725	January 29, 2019
Help needed with docker for reproducible research paper in knitr General Q&A	4	1656	February 22, 2015
Docker and general scientific computing set-ups	8	6147	July 27, 2015
Creating a package to reproduce an academic paper General Q&A package	23	3939	January 25, 2019
Packages Recommendation for Ubuntu 1604 Package Use Questions geospatial	8	2189	June 27, 2017

Tips for installing CUDA into a rocker docker container

Related topics