sutdDFL — Decentralised Federated Learning on NVIDIA Jetson Orin Nano

A research project by the Temasek Laboratories @ SUTD exploring Decentralised Federated Learning (DFL) deployed physically across a cluster of NVIDIA Jetson Orin Nano edge devices.

Full setup instructions, hardware configuration, and implementation details are documented in the Wiki.

---

Overview

Unlike centralised Federated Learning, this project implements peer-to-peer model aggregation across edge devices — no central server required. Each Jetson node trains locally on its own data partition and selectively exchanges model updates with neighbours based on the OCD-FL (Opportunistic Communication-efficient Decentralised FL) knowledge gain framework.

Key properties:

• No single point of failure — fully decentralised topology
• Privacy-preserving — only model weights are transmitted, never raw data
• Resource-aware — peer selection accounts for data distribution differences (Earth Mover's Distance) and device computation costs
• Edge-native — designed for ARM64 Jetson hardware with CUDA acceleration

---

Hardware & Environment

Component	Spec
Device	NVIDIA Jetson Orin Nano Dev Kit
JetPack	6.2.1
CUDA	12.6
OS	Ubuntu 22.04.5 LTS
Storage	128GB A2 microSD (U3/V30, A1/A2)

---

Repository Structure

sutdDFL/
└── ocdFL/          # OCD-FL algorithm implementation (see ocdFL/README.md)

---

Docker Image

The project runs inside a Docker container (sutd-dfl-jetson:v1) built on NVIDIA's JetPack-optimised PyTorch base image. The image is publicly available on GitHub Container Registry — no login required.

Pull the image on any Jetson:

docker pull ghcr.io/ngzhankang/sutd-dfl-jetson:v1

Run the container:

sudo docker run --runtime nvidia --net=host -v /home/$USER/SUTD:/app sutd-dfl-jetson:v1

For full Docker setup instructions including microSD configuration and NVIDIA Container Toolkit installation, see the Getting Started wiki page.

Updating the image (from any Jetson with the updated code):

# 1. Rebuild with a new version tag
sudo docker build --network=host -t sutd-dfl-jetson:v2 .

# 2. Tag for GHCR
sudo docker tag sutd-dfl-jetson:v2 ghcr.io/ngzhankang/sutd-dfl-jetson:v2

# 3. Login and push (only the pushing Jetson needs GITHUB_PAT in ~/.bashrc)
echo $GITHUB_PAT | sudo docker login ghcr.io -u ngzhankang --password-stdin
sudo docker push ghcr.io/ngzhankang/sutd-dfl-jetson:v2

# 4. Other Jetsons pull the new version (no login needed, image is public)
sudo docker pull ghcr.io/ngzhankang/sutd-dfl-jetson:v2

Bump the version tag (v2, v3 etc.) with each update — avoid reusing old tags so you always know what version is running on each device.

---

Implementations

See the Implementations wiki page for details on experiments and algorithm variants.

For the OCD-FL specific implementation, refer to ocdFL/README.md.

---

Team

Name	Role
Prof. Marie Therese Siew	Principal Investigator
Lucas Liew	Researcher
Skylar	Researcher
Ng Zhan Kang	Researcher

SUTD 2025