Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Towards Practical and Usable In-network Classification

Di Zhu, Jianxi Chen, Hyojoon Kim

TL;DR

ACORN tackles the practicality gap in in-network ML by automating the end-to-end deployment of Python ML models on distributed, programmable data planes. It introduces a layer-based representation for tree-based models and a streamlined SVM deployment, enabling large feature support and runtime programmability across multiple switches via an ILP-driven optimizer. The system demonstrates substantial feature capacity and latency benefits, achieving 65-90% faster per-request serving with negligible network overhead while supporting runtime updates without reloading devices. The authors provide a complete implementation (controller, translator, optimizer, and data-plane programs) and publicly release these artifacts, underscoring ACORN's potential for real-world deployment in programmable networks.

Abstract

In-network machine learning enables real-time classification directly on network hardware, offering consistently low inference latency. However, current solutions are limited by strict hardware constraints, scarce on-device resources, and poor usability, making them impractical for ML developers and cloud operators. To this end, we propose ACORN, an end-to-end system that automates the distributed deployment of practical machine learning models across the network. ACORN provides a fully automated pipeline that loads and deploys Python ML models on network devices using an optimized deployment plan from an ILP planner. To support larger models under hardware constraints and allow runtime programmability, ACORN adopts a novel data plane representation for Decision Tree, Random Forest, and Support Vector Machine models. We implement ACORN prototype in P4 and run it on real programmable hardware. Our evaluation shows ACORN can deploy classification ML models with 2-4x more features than state-of-the-art solutions, while imposing negligible overhead on network performance and traffic. We will make our data plane program, model translator, optimizer, and all related scripts publicly available.

Towards Practical and Usable In-network Classification

TL;DR

ACORN tackles the practicality gap in in-network ML by automating the end-to-end deployment of Python ML models on distributed, programmable data planes. It introduces a layer-based representation for tree-based models and a streamlined SVM deployment, enabling large feature support and runtime programmability across multiple switches via an ILP-driven optimizer. The system demonstrates substantial feature capacity and latency benefits, achieving 65-90% faster per-request serving with negligible network overhead while supporting runtime updates without reloading devices. The authors provide a complete implementation (controller, translator, optimizer, and data-plane programs) and publicly release these artifacts, underscoring ACORN's potential for real-world deployment in programmable networks.

Abstract

In-network machine learning enables real-time classification directly on network hardware, offering consistently low inference latency. However, current solutions are limited by strict hardware constraints, scarce on-device resources, and poor usability, making them impractical for ML developers and cloud operators. To this end, we propose ACORN, an end-to-end system that automates the distributed deployment of practical machine learning models across the network. ACORN provides a fully automated pipeline that loads and deploys Python ML models on network devices using an optimized deployment plan from an ILP planner. To support larger models under hardware constraints and allow runtime programmability, ACORN adopts a novel data plane representation for Decision Tree, Random Forest, and Support Vector Machine models. We implement ACORN prototype in P4 and run it on real programmable hardware. Our evaluation shows ACORN can deploy classification ML models with 2-4x more features than state-of-the-art solutions, while imposing negligible overhead on network performance and traffic. We will make our data plane program, model translator, optimizer, and all related scripts publicly available.

Paper Structure

This paper contains 34 sections, 8 equations, 12 figures, 9 tables.

Table of Contents

  1. Introduction
  2. Background and Motivation
  3. In-network computing and ML
  4. The Gaps in Distributed In-network ML
  5. ACORN: Design Principles and Workflow
  6. Design Principles and Key Techniques
  7. Workflow
  8. Model Representation and Translation
  9. Decision Trees
  10. Random Forest
  11. Support Vector Machine
  12. Optimization
  13. Assumption and Problem Definition
  14. Optimization Objectives
  15. Constraints
  16. ...and 19 more sections

Figures (12)

  • Figure 1: An overview of ACORN's workflow.
  • Figure 2: Distributing a decision tree (Iris dataset iris_53).
  • Figure 3: Encoding a decision tree in ACORN. On the tables, packet matches on the first three columns (black). Red columns are the actions.
  • Figure 4: Encoding a support vector machine model.
  • Figure 5: The data plane design of ACORN.
  • ...and 7 more figures