ARES accelerates training in resource-constrained devices and minimizes the effect of stragglers on the training through device-targeted split points while accounting for time-varying network throughput and computing resources. ARES takes into account application constraints (using energy sensitivity policy) to mitigate training optimization tradeoffs in terms of energy consumption and training time.

Python 3.7+ with Pytorch and torchvision for IoT devices (NVIDIA Jetson Nano and Raspberry Pi) and Server. Some pre-built PyTorch and torchvision pip wheel:

• Jetson Nano: https://forums.developer.nvidia.com/t/pytorch-for-jetson-version-1-8-0-now-available/72048
• Pytorch (Pi): https://discuss.pytorch.org/t/installing-pytorch-on-raspberry-pi-3/25215/12

• socket
• struct
• pickle
• numpy
• threading
• time
• matplotlib
• torch
• torchvision
• sys
• csv

CIFAR10 datasets can be downloaded manually and put into the datasets/CIFAR10 folder.

• CIFAR10: https://www.cs.toronto.edu/~kriz/cifar.html
• Altenatively, set the download parameter to True (on initial run) in torchvision.datasets.CIFAR10 in functions.py

To set the edge server and IoT devices configurations, edit the file configurations.py with the corresponding IP addresses, hostnames and port numbers for all devices and server. The default port number is set to 43000 for all devices. The configuration file is shared on all devices and server.

Launch ARES training on the server in ARES_training folder:

python3 ARES_Server.py --split True 

Launch ARES training on the IoT devices:

python3 ARES_Device.py --split True #requires sudo on Jetson Nano

The --split argument specifies an option to process all layers of the neural network model on the devices:

python3 ARES_Server.py --split False 
python3 ARES_Device.py --split False 

Run test ARES optimisation in ARESopt folder:

python3 ARES_optimiser.py

This requires all packages to installed.

The nvpmodel.conf file in Nano_power_modes folder resource configurations (power modes) for the Jetson Nano and can be manipulated to create custom ones. Update the change in /etc/nvpmodel.conf file.

To check the current power mode run:

sudo /usr/sbin/nvpmodel -q

To switch to another mode run:

sudo /usr/sbin/nvpmodel -m [MODE NUMBER] #the MODE NUMBER corresponds to the position of the target mode starting from 0

Note: the jtop interface for monitoring resources on the Jetson Nano sometimes does not run properly for custom power modes.

power monitoring in I2C folder:

/sys/bus/i2c/drivers/ina3221x/6-0040/iio:device0/in_power0_input

Modify the traffic_shapper.sh file in Network_traffic folder with the corresponding IP addresses and bandwidth limits. The qdiscs can be set for different network configurations accordingly.

To modify network traffic run:

sudo sh ./traffic_shapper.sh

To clear the qdisc on the device run: To modify network traffic run:

sudo sh ./clean_traffic_shapper.sh

Paper: https://doi.org/10.1016/j.comnet.2022.109380

Please cite the paper as follows: Samikwa, Eric, Antonio Di Maio, and Torsten Braun. "ARES: Adaptive Resource-Aware Split Learning for Internet of Things." Computer Networks 218 (2022): 109380.

@article{samikwa2022ares,
  title={ARES: Adaptive Resource-Aware Split Learning for Internet of Things},
  author={Samikwa, Eric and Di Maio, Antonio and Braun, Torsten},
  journal={Computer Networks},
  volume={218},
  pages={109380},
  year={2022},
  publisher={Elsevier}
}