Overview

EdgeTB is a hybrid testbed for distributed machine learning at the edge. It allows using Docker containers and physical nodes to create hybrid test environments. On the one hand, the existence of physical nodes improves the computing fidelity and network fidelity of EdgeTB, making it close to the physical testbed. On the other hand, compared with the physical testbed, the adoption of emulators makes it easier for EdgeTB to generate large-scale and network-flexible test environments.

Installation

1. At least 2 computing devices, one acts as Controller, and others act as Workers (as physical nodes, class: PhysicalNode or as emulator, class:Emulator).
2. Software requirement for computing devices.

Computing devices	Requirement
Controller	python3, python3-pip, NFS-Server
Worker (PhysicalNode)	python3, python3-pip, NFS-Client, iproute (iproute2)
Worker (Emulator)	python3, python3-pip, NFS-Client, iproute (iproute2), Docker

3. Copy controller into Controller and install the python packages defined in controller/ctl_req.txt.
4. Copy worker into Worker and install the python packages defined in worker/agent_req.txt.

File structure

controller
   ├─ dml_app  >>  Where we prepare roles, static files, shared by NFS
      ├─ dml_req.txt  >>  Role's execution environment
      ├─ Dockerfile  >>  Role's execution environment
      ├─ gl_peer.py  >>  Role's functions, an example
      ├─ nns  >>  Neural networks
      ├─ dml_utils.py
      └─ worker_utils.py
   ├─ dml_file  >>  Dynamically generated files for each node, transmitted over the network
      ├─ conf  >>  Generated by dml_tool/*_conf.py before running test, send to each node
      └─ log  >>  Received from each node
   ├─ dataset  >>  Splitted dataset, static files, shared by NFS
   ├─ dml_tool
      ├─ gl_dataset.json  >>  Dataset definition of all Gossip peer nodes, an example
      ├─ gl_structure.json  >>  Structure definition of all Gossip peer nodes, an example
      ├─ dataset_conf.py  >>  Used to generate dataset conf file for each node
      ├─ gl_structure_conf.py  >>  Used to generate structure conf file for each Gossip peer  node
      ├─ conf_utils.py
      ├─ splitter_utils.py
      └─ splitter_fashion_mnist.py  >>  Used to download and/or split dataset, an example
   ├─ gl_manager.py  >>  Runtime manager, an example
   ├─ gl_run.py  >>  Test environment definition, an example
   └─ links.json  >>  Network links definition
   
worker
   ├─ agent.py  >>  Used to communicate with controller/*_run.py
   ├─ dml_app  >>  mount point of controller/dml_app, over NFS
   ├─ dml_file
      ├─ conf  >>  Received from controller
      └─ log  >>  Generated by each node while running test, send to controller
   └─ dataset  >>  mount point of controller/dataset, over NFS

Usage

Workflow overview

Prepare roles, neural networks, dataset >> Define test environment >> Run it >> Collect result.

Workflow detail

1. The only things you need to do in Worker is to run worker/agent.py with python3 with root privileges. We need to mount NFS and install python packages via python3-pip, which require root privileges.
2. All the following operations should be completed in the Controller.
3. Prepare roles, just like what controller/dml_app/gl_peer.py does.
4. Prepare neural network model, just like what controller/dml_app/nns/nn_fashion_mnist.py does.
5. Prepare datasets and split it, just like what controller/dml_tool/nn_fashion_mnist.py does.
6. Update controller/dml_app/Dockerfile and controller/dml_app/dml_req.txt to meet your DML.
7. Prepare test environment controller/run.py, just like what controller/gl_run.py does.
8. Prepare network controller/links.json.
9. Prepare Runtime Manager, just like what controller/gl_manager.py does.
10. Run controller/run.py with python3 with root privileges and keep it running on a terminal (called Term).
11. It takes a while to deploy the tc settings, so please set your DML to start running after receiving a certain message, such as receiving a GET request for /start.
12. Wait until Term displays tc finish, and then start your DML.
13. Clear the test environment.

Examples

Gossip Learning

1. Same with above 1-4.
2. Just use the controller/dml_app/gl_peer.py, controller/dml_app/Dockerfile, and controller/dml_app/dml_req.txt
3. Modify controller/gl_run.py to define the test environment.
4. Modify controller/linlks.json to define the network.
5. Modify controller/dml_tool/gl_dataset.json to define the data used by each node and modify controller/dml_tool/gl_structure.json to define the DML structure of each node, see controller/dml_tool/README.md for more.
6. Run controller/gl_run.py with python3 with root privileges and keep it running on a terminal (called Term).
7. In path controller/dml_tool, type python3 dataset_conf.py -d gl_dataset.json in terminal to generate dataset conf files and type python3 gl_structure_conf.py -s gl_structure.json generate DML structure conf files.
8. Type curl localhost:3333/conf/dataset in a terminal to send those dataset conf files to each node. Wait until all nodes have received the dataset conf file. This function is defined in controller/base/manager.py.
9. Type curl localhost:3333/conf/structure to send those DML structure conf files to each node. Wait until all nodes have received the structure conf file. This function is defined in controller/base/manager.py.
10. Wait until Term displays tc finish.
11. Type curl localhost:3333/start in a terminal to start all nodes. This function is defined in controller/base/manager.py and controller/gl_manager.py.
12. When there is no node Gossip, type curl localhost:3333/finish in a terminal to stop all nodes and collect result files. This function is defined in controller/base/manager.py and controller/gl_manager.py.
13. Commands such as curl localhost:3333/emulated/reset and curl localhost:3333/physical/resetare used to remove all the emulated nodes and physical nodes. These functions are defined in controller/base/manager, and worker/agent.py.

Federated Learning

1. Same with above 1-4.
2. Just use the controller/dml_app/fl_trainer.py, controller/dml_app/fl_aggregator.py, controller/dml_app/Dockerfile and controller/dml_app/dml_req.txt
3. Modify controller/fl_run.py to define the test environment.
4. Modify controller/linlks.json to define the network.
5. Modify controller/dml_tool/fl_dataset.json to define the data used by each node and modify controller/dml_tool/fl_structure.json to define the DML structure of each node, see controller/dml_tool/README.md for more.
6. Run controller/fl_run.py with python3 with root privileges and keep it running on a terminal (called Term).
7. In path controller/dml_tool, type python3 dataset_conf.py -d fl_dataset.json in terminal to generate dataset conf files and type python3 fl_structure_conf.py -s fl_structure.json generate DML structure conf files.
8. Type curl localhost:3333/conf/dataset in a terminal to send those dataset conf files to each node. Wait until all nodes have received the dataset conf file. This function is defined in controller/base/manager.py.
9. Type curl localhost:3333/conf/structure to send those DML structure conf files to each node. Wait until all nodes have received the structure conf file. This function is defined in controller/base/manager.py.
10. Wait until Term displays tc finish.
11. Type curl localhost:3333/start?root=n1 in a terminal to start all nodes. This function is defined in controller/base/manager.py and controller/fl_manager.py. The root should be the first node defined in controller/dml_tool/fl_structure.json.
12. When the pre-set training round is met, it will automatically stop all nodes and collect result files. This function is defined in controller/base/manager.py and controller/fl_manager.py.
13. Commands such as curl localhost:3333/emulated/reset and curl localhost:3333/physical/resetare used to remove all the emulated nodes and physical nodes. These functions are defined in controller/base/manager, and worker/agent.py.

E-Tree Learning

1. Same with above 1-4.
2. Just use the controller/dml_app/el_peer.py, controller/dml_app/Dockerfile, and controller/dml_app/dml_req.txt
3. Modify controller/el_run.py to define the test environment.
4. Modify controller/linlks.json to define the network.
5. Modify controller/dml_tool/el_dataset.json to define the data used by each node and modify controller/dml_tool/el_structure.json to define the DML structure of each node, see controller/dml_tool/README.md for more.
6. Run controller/el_run.py with python3 with root privileges and keep it running on a terminal (called Term).
7. In path controller/dml_tool, type python3 dataset_conf.py -d el_dataset.json in terminal to generate dataset conf files and type python3 el_structure_conf.py -s el_structure.json generate DML structure conf files.
8. Type curl localhost:3333/conf/dataset in a terminal to send those dataset conf files to each node. Wait until all nodes have received the dataset conf file. This function is defined in controller/base/manager.py.
9. Type curl localhost:3333/conf/structure to send those DML structure conf files to each node. Wait until all nodes have received the structure conf file. This function is defined in controller/base/manager.py.
10. Wait until Term displays tc finish.
11. Type curl localhost:3333/start?root=n1 in a terminal to start all nodes. This function is defined in controller/base/manager.py and controller/el_manager.py. The root should be the first node defined in controller/dml_tool/el_structure.json.
12. When the pre-set training round is met, it will automatically stop all nodes and collect result files. This function is defined in controller/base/manager.py and controller/el_manager.py.
13. Commands such as curl localhost:3333/emulated/reset and curl localhost:3333/physical/resetare used to remove all the emulated nodes and physical nodes. These functions are defined in controller/base/manager, and worker/agent.py.

Ring All-Reduce

1. Same with above 1-4.
2. Just use the controller/dml_app/ra_peer.py, controller/dml_app/Dockerfile, and controller/dml_app/dml_req.txt
3. Modify controller/ra_run.py to define the test environment.
4. Modify controller/linlks.json to define the network.
5. Modify controller/dml_tool/ra_dataset.json to define the data used by each node and modify controller/dml_tool/ra_structure.json to define the DML structure of each node, see controller/dml_tool/README.md for more.
6. Run controller/ra_run.py with python3 with root privileges and keep it running on a terminal (called Term).
7. In path controller/dml_tool, type python3 dataset_conf.py -d ra_dataset.json in terminal to generate dataset conf files and type python3 ra_structure_conf.py -s ra_structure.json to generate DML structure conf files.
8. Type curl localhost:3333/conf/dataset in a terminal to send those dataset conf files to each node. Wait until all nodes have received the dataset conf file. This function is defined in controller/base/manager.py.
9. Type curl localhost:3333/conf/structure to send those DML structure conf files to each node. Wait until all nodes have received the structure conf file. This function is defined in controller/base/manager.py.
10. Wait until Term displays tc finish.
11. Type curl localhost:3333/start in a terminal to start all nodes. This function is defined in controller/base/manager.py and controller/ra_manager.py.
12. When the pre-set training round is met, it will automatically stop all nodes and collect result files. This function is defined in controller/base/manager.py and controller/ra_manager.py.
13. Commands such as curl localhost:3333/emulated/reset and curl localhost:3333/physical/resetare used to remove all the emulated nodes and physical nodes. These functions are defined in controller/base/manager, and worker/agent.py.

Citing

Please cite our paper if you find EdgeTB is useful in your research.
Lei Yang, Fulin Wen, Jiannong Cao, Zhenyu Wang. "EdgeTB: a Hybrid Testbed for Distributed Machine Learning at the Edge with High Fidelity." IEEE Transactions on Parallel and Distributed Systems. DOI: 10.1109/TPDS.2022.3144994.

Contact

EdgeTB is designed and developed by the joint research team at School of Software Engineering, South China University of Technology, and the Department of Computing, The Hong Kong Polytechnic University. If you have any question, please contact with us: Fulin Wen [email protected] and Lei Yang [email protected].