Docker containers for PYME's distributed servers - #1219

nvidia-container-toolkit (https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/overview.html) must be installed on the Docker host machine for this to work. Launch the GPU-enabled Docker server (main node uses --main flag, other nodes do not have this flag) with

docker run --rm -dp 9000:9000 --gpus all pyme-server --main

Looks like we might also be able to run the multi-process service. Wild. Would be neat from a performance standpoint, but if its a headache its a headache
NVIDIA/nvidia-docker#419

Also big thanks for doing this Zach!

Would be cool. And no problem. Once this is merged/built on our own Docker Hub account (presumably python-microscopy? or maybe david_baddeley for continuity?) we can write up the instructions for install, which will be something like.

Before you begin

Note that this could create some big security holes in your network. We recommend running this on a separate, private network connected only to your microscope computer.

Requirements

• GNU/Linux x86_64 with kernel version > 3.10
• Docker >= 19.03
• NVIDIA GPU with Architecture >= Kepler (or compute capability 3.0)
• NVIDIA Linux drivers

Installation

1. Install NVIDIA container toolkit (https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/overview.html).
2. Run docker pull python-microscopy/pyme-server:latest.

Run

1. Enter docker run --rm -d --gpus all pyme-server --main to launch the main node of a PYME server array.
2. For each additional machine you wish to bring into the server array, enter docker run --rm -d --gpus all pyme-server.

NOTE: There is currently no python manage.py runserver 9000 call (https://python-microscopy.org/doc/cluster/cluster_install.html#cluster-install). Maybe we want to add this?