Skip to content

Latest commit

 

History

History
95 lines (72 loc) · 3.56 KB

README.md

File metadata and controls

95 lines (72 loc) · 3.56 KB
Raw

BGPsyche

BGPsyche is a system to predict BGP paths between arbitrary autonomous systems (ASes) using deep learning. It was written as part of my (@dominiksta) masters thesis. It should be considered a prototype implementation, not a finished product. Nevertheless, running (and training) BGPsyche should be relatively straightforward and is outlined in this README.

BGPsyche is built for both PyPy and CPython 3.10 on Ubuntu 22.04. Other OS/Python combinations may or may not work.

System Dependencies

  • PyPy 3.10 and CPython 3.10
  • A Rust toolchain: BGPsyche uses the pybgpkit-parser MRT parser, which is written in Rust. The download and compilation of this module will be done automatically by pip.

Install and Run

Start by creating an env.ini file in /bgspyche, following the template env.ini.template. You will have to create accounts for both PeeringDB and MaxMind GeoLite2. Then run the following commands:

pypy3.10 -m venv .venv-pypy
. .venv-pypy/bin/activate
pip install -r bgpsyche/requirements.pypy.txt
deactivate

python3.10 -m venv .venv
. .venv/bin/activate
pip install -r bgpsyche/requirements.main.txt

# this can take up to half an hour on the initial sync
python -m bgpsyche peeringdb_sync

python -m bgpsyche --help

Usage

TODO

When BGPsyche runs for the first time, it will have to pull down quite a few datasets from the internet and compute various features from those. Please just be patient here. There may also be cases where you will run out of memory on the first run and have to just run the same command again. As stated, its a prototype implementation. All datasets and computed features are cached on disk though, so re-running the same command will be much quicker and should not cause any issues.

Using the pre-trained ("silver") model

BGPsyche offers a pre-trained model in the repository. This model was trained with what was found to be the ideal parameters and features in the thesis. It was therefore called the "silver" model.

Usage of a pre-trained model requires about 6 GB of RAM. Almost all of the memory is taken up by various datastructures, not the model itself. Running a larger model should therefore take a comparable amount of memory.

TODO date of data (not just 23-05-1)

BGPsyche exposes an HTTP API with a single endpoint for path predictions:

# in one terminal
python -m bgpsyche listen 8080
# in another terminal
curl "http://localhost:8080/predict?source=3320&sink=6939"

Training

Training a model with default configuration (another "silver" model) takes about 16GB of RAM and 16GB of VRAM. Please see make_dataset.py for adjusting the dataset size if necessary. In general, if you want to train your own model you probably want to read (or at least skim) the thesis. The most relevant things to adjust are arguably in vectorize_features.py, enrich.py, make_dataset.py and classifier_nn.py.

# to evaluate the model by comparing to real paths 
# (with output tensorboard for visualization)
python -m bgpsyche train_and_evaluate

# in another terminal
python -m bgpsyche tensorboard