Toward In-Network Intelligence: Running Distributed Artificial Neural Networks in the Data Plane

• Published in: IEEE communications letters Vol. 25; no. 11; pp. 3551 - 3555
• Main Authors: Saquetti, Mateus, Canofre, Ronaldo, Lorenzon, Arthur F., Rossi, Fabio D., Azambuja, Jose Rodrigo, Cordeiro, Weverton, Luizelli, Marcelo C.
• Format: Journal Article
• Language: English
• Published: New York IEEE 01.11.2021; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Artificial neural networks; Biological neural networks; Data models; Distributed databases; In-band telemetry; In-network intelligence; Intelligence; Neural networks; Neurons; Optimization; Programmable forwarding planes; Provisioning; Real-time systems; Switches; Telemetry; Visibility
• ISSN: 1089-7798; 1558-2558
• DOI: 10.1109/LCOMM.2021.3108940

In this letter, we make a case for in-network intelligence in programmable data planes (PDPs) by taking the first steps toward running distributed Artificial Neural Networks (ANNs) in programmable switches. The main novelty of our research lies in distributing the neurons of an ANN into multiple switches instead of running an entire ANN in a single device. The many advantages of this approach include wider network flow visibility and better resource usage across switches and links. We discuss the research challenges involved in expressing neuron logic for PDPs, mapping neurons to switches, and enabling neuron communication. To tackle these challenges, we introduce PDP programming constructs for performing neuron computation, formalize an optimization model for neuron placement, and tailor in-band telemetry for neuron inter-communication using production flows. Results obtained with a P4 implementation evidence that our approach improves network management tasks while keeping their provisioning overhead similar to a baseline.