[論文レビュー] IIsy: Practical In-Network Classification

Changgang Zheng, Zhaoqi Xiong+7 more|arXiv (Cornell University)|May 17, 2022

Molecular Communication and Nanonetworks被引用数 33

ひとこと要約

IIsyは、従来型とアンサンブルML分類器の範囲を市販のプログラム可能スイッチへマッピングし、ネットワーク内での分類を可能にします。小規模なネットワーク内モデルとより大きなバックエンドモデルを組み合わせたハイブリッドモデルを含みます。

ABSTRACT

The rat race between user-generated data and data-processing systems is currently won by data. The increased use of machine learning leads to further increase in processing requirements, while data volume keeps growing. To win the race, machine learning needs to be applied to the data as it goes through the network. In-network classification of data can reduce the load on servers, reduce response time and increase scalability. In this paper, we introduce IIsy, implementing machine learning classification models in a hybrid fashion using off-the-shelf network devices. IIsy targets three main challenges of in-network classification: (i) mapping classification models to network devices (ii) extracting the required features and (iii) addressing resource and functionality constraints. IIsy supports a range of traditional and ensemble machine learning models, scaling independently of the number of stages in a switch pipeline. Moreover, we demonstrate the use of IIsy for hybrid classification, where a small model is implemented on a switch and a large model at the backend, achieving near optimal classification results, while significantly reducing latency and load on the servers.

研究の動機と目的

• Motivate and enable machine learning classification directly in the network to reduce server load and latency.
• Provide a mapping framework that translates trained ML models into data-plane and control-plane programs for commodity switches.
• Support a range of models (decision trees, Random Forest, Isolation Forest, XGBoost, SVM, Naïve Bayes, K-Means) while avoiding neural networks.
• Address resource constraints and feature extraction challenges in the network to make in-network classification feasible.
• Demonstrate the efficacy of hybrid deployments that split computation between the switch and backend servers.

提案手法

• Develop a model-to-switch mapping methodology that uses lookup tables to implement math operations and classifications.
• Decouple pipeline depth from model depth so that the number of switch stages does not grow with tree depth.
• Share feature tables across trees in ensemble methods to reduce table proliferation.
• Provide two approaches for classical models (SVM, Naïve Bayes, K-Means): table-per-feature or table-per-class/indicator to balance accuracy and resource use.
• Enable retraining and updates via table updates without changing the deployed program, including hybrid deployment workflows.
• Autogenerate data-plane (P4) and control-plane table entries from standard ML training outputs (e.g., pickle) for target devices.

実験結果

リサーチクエスチョン

• RQ1 How can a variety of ML classification models be mapped efficiently onto programmable network devices without altering network protocols or tools?
• RQ2 What mechanisms enable scalable in-network classification that decouples model complexity from switch pipeline depth?
• RQ3 How can feature extraction be performed across packet, flow, aggregate, and file granularity within network devices?
• RQ4 Can a hybrid deployment (small in-network model + large backend model) achieve near-optimal accuracy with reduced backend load and latency?
• RQ5 What are practical retraining and update workflows for in-network models that minimize traffic disruption?

主な発見

• IIsy demonstrates mapping of several ML classifiers to programmable switches, including decision trees, Random Forest, Isolation Forest, XGBoost, SVM, Naïve Bayes, and K-Means.
• A mapping approach that is independent of the number of pipeline stages enables scaling of ensemble models on switches.
• Feature extraction is demonstrated across packet, flow, aggregate, and file granularity, including handling of data stored across packets.
• Hybrid deployments show reduced backend load and lower classification latency while preserving classifier performance for time-sensitive tasks.
• Retraining can be done through table updates without changing the deployed data-plane program, enabling quick model updates in production.