This repository contains scripts and instructions for reproducing the experiments in our OSDI '21 paper "Oort: Efficient Federated Learning via Guided Participant Selection".
Our install.sh
will install the following automatically:
- Anaconda Package Manager
- CUDA 10.2
Note: if you prefer different versions of conda and CUDA, please check comments in install.sh
for details.
Run the following commands to install Oort.
git clone https://github.com/SymbioticLab/Oort
cd Oort
source install.sh
The output of the experiment will validate the following major claims in our evaluation (section 7 in paper):
- Oort outperforms existing random participant selection by 1.2×-14.1× in time-to-accuracy performance, while achieving 1.3%-9.8% better final model accuracy (§7.2.1) -> Table 1 and Figure 9.
- Oort achieves close-to-optimal model efficiency by adaptively striking the trade-off between statistical and system efficiency with different components (§7.2.2) -> Figure 11 and 12.
- Oort outperforms its counterpart over a wide range of parameters and different scales of experiments, while being robust to outliers (§7.2.3) -> Figure 13, 14 and 15.
- Oort can serve developer testing criteria on data deviation while reducing costs by bounding the number of participants needed even without individual data characteristics(§7.3.1) —> Figure 16.
- With the individual information, Oort improves the testing duration by 4.7× w.r.t. Mixed Integer Linear Programming (MILP) solver, and is able to efficiently enforce developer preferences across millions of clients (§7.3.2) -> Figure 17.
Please go to ./training
directory and follow the training README to run training scripts.
Please go to ./testing
directory and follow the testing README to run testing scripts.
Repo Root
|---- Oort # Oort code base.
|---- training
|---- evals # Submit/terminate training jobs
|---- configs # Configuration examples
|---- testing # Testing scripts
please consider to cite our paper if you use the code or data in your research project.
@inproceedings{Oort-osdi21,
title={Efficient Federated Learning via Guided Participant Selection},
author={Fan Lai and Xiangfeng Zhu and Harsha V. Madhyastha and Mosharaf Chowdhury},
booktitle={USENIX Symposium on Operating Systems Design and Implementation (OSDI)},
year={2021}
}
Thanks to Qihua Zhou for his Falcon repo.
Fan Lai (fanlai@umich.edu) and Xiangfeng Zhu (xzhu0027@gmail.com)