GitLab

CUDA 11 Fix

- use one of the pynvml packages based on compatible Python version
- fix a change in API for clock speed
- convert device type and output from subprocess from bytes to string

In order to allow development and not require to publish to PyPI before building
the docker image.

Since it needs the NVML runtime it's based on CUDA image and needs
nvidia-docker. The whole CUDA is not needed, so we could possibly optimize it.

It needs the compiler to build the psutil Python package.
We could use some build image and not put the compiler in the target image.

We mount the master's config file from host into the container.

The agent hostname in Docker is meaningless (some random number),
but the URL (or "self") might be more descriptive.

It's not perfect - we need to provide path to the webapp.py.

It's been superseded with the distibuted web app approach.

- only nvgpu.webapp
- master selected by config file via NVGPU_CLUSTER_CFG
  with non-empty AGENTS list
- query from master to master is done directly, not via HTTP
  (to prevent nested request on single-threaded app)
- make the agents configurable

- agents are Flask apps with REST API returning machine GPU status in JSON
- master is a Flask app that gathers statuses from agents and shows aggregate
  overview as an HTML page

It synchronously gathers output from `nvl` from local and remote machines
via SSH and formats the colored ANSI output via ansi2html.

Function has toe be separated from module by colon, not dot.

Replace with empty string.