Repositories for computing optimal placement and running partitioned DNNs.
The Wedge framework includes two main components:
- The Wedge workers (
wedge-woker), which include an ONNX runtime, synchronization mechanisms, physical network monitoring and I/O management with other workers and a potential orchestrator. - The Wedge orchestrator (
wedge-orchestrator) manages the deployed workers, extract configuration from the DNN, and compute optimal placements based on physical network information retrieved from the workers. The orchestrator also includes an API to interact with the Wedge service.
This repository provides both docker-compose and kubernetes support and configuration to run examples of the Wedge framewok.
To run sample deployments on docker-compose or kubernetes, follow the steps described in the respective folders.
- DNN Partitioning for Inference Throughput Acceleration at the Edge, T. Feltin, L. Marché, J-A. Cordero-Fuertes, F. Brockners and T. H. Clausen, in IEEE Access, vol. 11, pp. 52236-52249, 2023, doi: 10.1109/ACCESS.2023.3244497