CluBoard

Simplistic web monitor of GPU utilization. Best suited for systems with multiple GPUs and many users. Built upon gpustat.

Idea of CluBoard is to have all the simple outline of the cluster GPUs availability at a single glance. Current features:

• nodes can be arranged in a grid with the layout defined by a user and given aliases
• shows all processes on each GPU, their owners and PIDs
• along with GPUs statistics, current CPU + RAM + Swap usage are reported for each node
• top users by GPU usage
• highlighting the users who allocated too many GPUs
• specification of the lab names

The monitor can be an additional tool for clusters equipped with job management systems such as SLURM, IBM Spectrum LSF, etc.

Installation

Please download the repo from the master branch and install the requirements. The commands below will do it for you.

git clone https://github.com/seva100/cluboard.git
cd cluboard
pip install -r requirements.txt

Usage

First, please choose a login node on your cluster from which all other nodes will be accessible.

As a first step, the config file config.yaml needs to be changed. Open it and specify the addresses or IPs of the servers in the server_names property. It is a 2D nested list that will correspond to the arrangement of nodes in the CluBoard web page. Please see the comments under each property in the config.yaml.

---

NB: make sure that all the nodes can be accessible by the names specified in the server_names field from the login node by SSH without password. To ensure that, type in the shell at the login node:

$ ssh one_of_my_servers.com

If everything is OK, the shell of the one_of_my_servers.com will open. If the password is being asked, make sure to copy the public key to the one_of_my_servers.com for the no-password access. On Ubuntu and similar systems, the command ssh-copy-id one_of_my_servers.com, executed from the login node, will do that for you.

---

Make sure that the gpustat pip package is installed on all machines to be accessed. If the machines do not share the same environment, running pip install gpustat on each of them will fix it.

To launch the monitor,

bash gpu_monitor_start.sh [-h] [-p|--port <port>] [-d|--debug] [-c|--config <config_path>]

Options:

• -h: print help message
• -p <port> or --port <port>: run monitor on port <port>. Default port: 8888.
• -d or --debug: debug mode. Logs from application will be written to console, and any change of .py files of the project will make it being reloaded automatically. Usually only needed when the files of the project are being modified.
• -c or --config: specify the config file if it's different that config.yaml.

We recommend to launch CluBoard in one of the session managers, e.g. screen or tmux, that will not interrupt the session at the logout.

---

To be done

• distribution of GPUs per lab
• groups labs by users not by nodes

PRs

Pull requests & suggestions are much appreciated — please don't hesitate to open an issue or to contribute. Since the codebase is in the very early state, bugs for various systems can be present -- please open an issue in this case as well.