bmon is a CLI tool for monitoring system metrics and diagnosing potential bottlenecks in GPU-accelerated training.
bmon has 4 features, which can be turned on/off mostly independently:
- Display key GPU metrics (utilization, memory usage, temperature, power etc.)
- Display key CPU metrics for running GPU compute processes (CPU utilization, RAM etc.)
- Display key disk metrics
- Infer potential problems and bottlenecks from metrics (e.g. suspected disk bottleneck, GPU thermal throttling etc.)
With most features turned off, bmon may be used as a minimalist nvidia-smi
. With all features enabled, it becomes your personal research assistant.
bmon is in the alpha stage of development. Until version 1.0, there may be bugs, missing features, and API changes between versions.
Different labs often have very different compute setups and training pipelines optimized for one setup are likely to not be optimal for others. Defaults in libraries and other people's code are often poor -- to get the most out of our resources, we should all be tuning our training pipelines for our own setups.
When I am prototyping and training deep learning models, I am constantly watching nvidia-smi
and htop
to get an idea of how well I am utilising my system resources. This works ok, but has a few downsides:
- visual clutter (both tools give large outputs, where only a few key metrics are actually useful for me).
- limited insight (these tools on their own do not monitor disk/network, I could use more tools like
iostat
, but that exacerbates problem 1). - interpretation effort (surely we should be able to automate the interpretation of system metrics to diagnose issues!)
bmon attempts to solve these problems by presenting a clean and simple way of monitoring resources. I am particularly interested in its potential to give researchers informed analysis about program performance.
The binaries for each version are uploaded to GitHub as 'releases', which can be viewed in the UI. Simply download the latest one and move to a location in your $PATH
.
wget https://github.com/Charl-AI/bmon/releases/download/<VERSION>/bmon
chmod +x ./bmon
mv ./bmon <YOUR_PREFERRED_DIR_IN_$PATH>
You may also build from source by cloning this repo and running cargo build --release
.
bmon builds on existing command line tools for system monitoring. Most linux machines with working NVIDIA GPUs should satisfy the requirements already. In practice, you'll be fine if you can run the following commands without errors: nvidia-smi
, free
, nproc
, iostat
, ps
.
Show options with bmon -h
.
Using as an nvidia-smi replacement: bmon
With all features: bmon --all
Tip: use the linux watch
command to refresh stats every n seconds (e.g. watch -n 5 bmon
)
Short term:
- improve error handling (allow us to recover from non-critical issues)
- add optional disk metrics monitoring from iostat output
- Write bottleneck diagnosis document explaining how to use the metrics to find problems in typical training pipelines
- Add bottleneck diagnoses from the document to the automatic bottneck inference feature
- Find a way to monitor network metrics
Long term:
- Allow for outputs other than just printing (e.g. to csv/json/sqlite)
- Log metrics so we can diagnose bottlenecks with a time-component (currently bmon simply displays a snapshot each time it is called and all the bottlneck diagnosis sees only the snapshot)
- Make the tool interactive? Adjust fan speed, kill processes etc using bmon
I would love some help with this project! I am mostly building this tool because I want to use it, but I would be even more happy if other people want to use it :) Suggestions, issues, and PRs are welcome.