Ocularone Scheduler

Ocularone Scheduler offers heuristics for scheduling of inferencing tasks over real-time drone videos for helping navigate Visually Impaired Persons (VIPs). It's real-time deadline-aware schedulier operates across edge accelerators and cloud. The heuristic algorithms try to maximize the Quality of Service (QoS) and Quality of Experience (QoE). These algorithms are validated using a containerized emulation-based testbed based upon UltraViolet. We also perform practical validation of these scheduling strategies on real hardware using edge accelerators and DJI Tello nano-drones and compare them against state-of-the-art baselines.

Attribution

If you use the Ocularone Scheduler, please cite our work!

"Adaptive Heuristics for Scheduling DNN Inferencing on Edge and Cloud for Personalized UAV Fleets", Suman Raj and Radhika Mittal and Harshil Gupta and Yogesh Simmhan, Future Generation Computer Systems, 2025, 10.1016/j.future.2025.107874.

@article{RAJ2025107874,
  title = {Adaptive heuristics for scheduling DNN inferencing on edge and cloud for personalized UAV fleets},
  journal = {Future Generation Computer Systems},
  pages = {107874},
  year = {2025},
  issn = {0167-739X},
  doi = {https://doi.org/10.1016/j.future.2025.107874},
  url = {https://www.sciencedirect.com/science/article/pii/S0167739X25001694},
  author = {Suman Raj and Radhika Mittal and Harshil Gupta and Yogesh Simmhan},
  keywords = {Edge and cloud computing, Unmanned aerial vehicles, Intelligent offloading, Deadline-aware scheduling, DNN inferencing}
}

System Requirements

To run experiments, you would need a machine with atleast 30 vCPUs, as each edge emulation container gets pinned to CPU cores. We use a host machine with an Intel Xeon 6208U [email protected] GHz with 30 vCPUs and 128 GB RAM, running Ubuntu v18.04.

Other requirements include (version tested on):

openjdk (version 11.0.19)
Apache Maven (version 11.0.19 from https://maven.apache.org/install.html)
Python 3.6.9
tmux (to help with terminal and session management)

UltraViolet

Environment Setup

UltraViolet Setup

Set up Ultraviolet as per instructions in https://github.com/dream-lab/UltraViolet/blob/master/README.md. You could do either automatic installation or manual installation from Step 1-6. Once it's installed, please follow the steps below:
Create three new terminal windows: one each for the frontend server, controller and worker
- In the first terminal window, start the frontend server.
```
sudo MaxiNetFrontendServer
```
- In the second terminal window, start the controller.
```
cd <Path to Your UltraViolet directory>/pox
python3 pox.py forwarding.l2_learning
```
- In the third terminal window, start the worker.
```
sudo MaxiNetWorker
```
Enter choice 1 for each IP choice when prompted.

These can be left running across experiment runs and need not be shut down or restarted each time.

Drone Containers Setup

Ensure you have a functional docker setup on your host machine. (We use Docker Version 20.10.8, build 3967b7d)
Pull the drone docker image from dockerhub using

sudo docker pull radmittal3/buddy_drone:warmup

Ensure that docker image radmittal3/buddy_drone:warmup is present on your host machine with the command

sudo docker images

radmittal3/buddy_drone:warmup must be one of the entries.

Tag the docker image with

sudo docker tag radmittal3/buddy_drone:warmup buddy_drone:warmup

as the Ultraviolet setup look for a docker image referenced as buddy_drone:warmup.

Edge Containers Setup

Build the edge base docker image using the <path to ocularone scheduler>/edge_files/Dockerfile_edge. Name and tag the image as edge_device:1.0.
Ensure that docker image edge_device:1.0 is present on your host machine with the command as the bringup files look for a docker image referenced as edge_device:1.0.
Create an edge container with the docker image and access its CLI using

sudo docker run -it edge_device:1.0 bash

In the edge container's CLI, copy the required contents such as the inferencing script, post processing scripts etc from the host machine with scp.

scp <host machine's username>@<host machine's IP address>:<path to ocularone scheduler>/edge_files/* .

In the edge container, ensure that executor.sh points to the files copied in the above step. If not, edit the paths. Check and edit the contents with

vi executor.sh

You may optionally use VS code's Dev Containers extension (https://code.visualstudio.com/docs/devcontainers/attach-container) to attach to and develop inside the container.

Once all changes within the edge container are complete, CHANGES MUST BE SAVED to your edge docker image by running

sudo docker commit <edge container ID> edge_device:1.0

sudo docker ps is useful to find the ID of running docker containers. docker ps -a may be used in case the container has been stopped.

AWS Lambda (Cloud) Functions Setup

Use the folders in <path to ocularone scheduler>/lambda for each DNN model. We provide the models for BODY_POSE_ESTIMATION, CROWD_DENSITY & MASK_DETECTION DNN models, which may be used for testing the scheduler.
Use AWS SAM CLI to deploy the models as AWS functions. (Check out https://aws.amazon.com/blogs/compute/using-container-image-support-for-aws-lambda-with-aws-sam/)
Create function URLs for each function from the AWS Lambda Console and tune parameters including memory allocated as per requirements (described in the experimental setup section of https://arxiv.org/abs/2412.20860v1).
Add the function URLs to lines 261-266 in <path to ocularone scheduler>/src/main/java/org/dreamlab/ApplicationDetails/Application.java and mock_edge.py in edge_files.