Dissertação de Mestrado Acadêmico de João Vicente Fatur Lessa

Defesa de Dissertação de Mestrado Acadêmico

Aluno(a): João Vicente Fatur Lessa
Orientador(a): Luciano Paschoal Gaspary

Título: In-Network Inference with NEURALP4: Auto-Generating Full-Fledged NN Models for PDPs
Linha de Pesquisa: Arquiteturas, Protocolos e Gerência de Redes e Serviço

Data: 18/11/2025
Hora: 10:30
Local: Esta banca ocorrerá de forma híbrida (virtual e presencial), na sala Sala 215, Prédio 43.412 do Instituto de Informática/UFRGS e pelo link https://mconf.ufrgs.br/webconf/00010356.

Banca Examinadora:
-Alberto Egon Schaeffer Filho (Universidade Federal do Rio Grande do Sul)
-Ronaldo Alves Ferreira (Universidade Federal do Mato Grosso do Sul)
-Weverton Luis da Costa Cordeiro (Universidade Federal do Rio Grande do Sul)

Presidente da Banca: Luciano Paschoal Gaspary

Resumo: The emergence of data plane programmability has opened new opportunities for embedding computation directly into the network. A prominent example is the offloading of Neural Network (NN) inference into programmable switches, where results can be produced at line rate and without reliance on external servers. Yet, deploying full-fledged NNs in the data plane remains challenging due to severe constraints on memory, arithmetic precision, and supported operations. As a result, most existing approaches resort to heavy simplifications such as binarization, often sacrificing accuracy and limiting applicability. In this master’s thesis, we introduce NEURALP4, an automated system for generating NN models that execute inference within programmable forwarding devices. NEURALP4 eliminates the need for manual coding by automatically producing all necessary switch code and configuration files, thereby lowering the barrier to deploying in-network inference. Our approach avoids aggressive simplifications by addressing device limitations through a labor-division strategy that distributes NN layers across multiple switches, together with customized algorithms based on numerical conversion and algebraic transformations. These mechanisms enable standard NNs to be mapped efficiently into restricted data plane environments. We implemented a proof-of-concept and evaluated NEURALP4 across five representative use cases. The results shows that the system preserves the accuracy of conventional server-based NNs while maintaining a memory footprint compatible with existing hardware constraints. Furthermore, the generated code requires minimal manual intervention, reducing the complexity and error-proneness of deploying in-network inference. This work contributes to the vision of intelligent networks by bridging high-level machine learning models with low-level programmable devices. It highlights how programmable switches can serve as platforms for distributed intelligence in applications of many areas. NEURALP4 thus contributes to bridging networking and AI, demonstrating the feasibility of in-network inference that extends beyond simplifications toward complete NN models.

Palavras-Chave: Software-Defined Networking. Programmable Data Planes. In-Network Computing. In-Network Inference. Neural Networks. Traffic Classification. Code Generation.