Learning-aided SDC Control with Programmable Switches

Abstract Software Defined Coalition (SDC) is a promising architecture to manage and control network resources and meet the requirements of easy configuration and agility to coalition needs in tactical environments. Machine learning techniques can aid SDC in automatically deriving optimal control policies. Natu- rally, the training and inference tasks of the learning models can be performed at the powerful SDC control plane, or controllers, each managing a different enclave in the coalition network. Yet, network dynamics can cause controllers to be fragmented from the nodes they manage rendering impossible their access to the learning models and thus their in-time resource reconfiguration. To address this, we propose an extension of SDC architecture with programmable switches capable of running locally (without the controllers) appropriately designed learning models. Our key idea is to use Binarized Neural Networks (BNNs) to significantly reduce the computation and memory requirement of the model to a level that a switch can afford. To train these models across different coalition enclaves, we adopt a Federated Learning approach. Evaluations on a prototype demonstrate that a multi- fold improvement in latency and communication overheads can be achieved compared to state-of-the-art learning architectures for packet classification as an example of a demanding SDC control application.
  • Qiaofeng Qin (Yale)
  • Konstantinos Poularakis (Yale)
  • Leandros Tassiulas (Yale)
  • Kin K. Leung (Imperial)
  • Franck Le (IBM US)
  • Paul Yu (ARL)
Date Sep-2020
Venue 4th Annual Fall Meeting of the DAIS ITA, 2020