Scholarlink

With the rapid proliferation of application scenarios such as the Internet of Things (IoT), industrial automation, smart homes, and vehicular networks, edge devices are facing increasing demands for low-latency and high-performance intelligent data processing. However, typical edge nodes—including home gateways, edge servers, and Customer Premises Equipment (CPE)—often suffer from limited computational resources, constrained memory capacity, and tight energy budgets. These constraints make it difficult for traditional user-space inference frameworks to achieve millisecond-level responsiveness under high-frequency data streams. To address this issue, this paper proposes a lightweight in-kernel Convolutional Neural Network (CNN) inference framework based on the extended Berkeley Packet Filter (eBPF), aiming to enable real-time traffic identification and classification on edge devices. By embedding the CNN model directly into the eBPF program of the Linux kernel, the proposed method performs packet parsing, feature extraction, and model inference entirely in kernel space, thereby eliminating the overhead of context switches and system calls common in user-space solutions. Experiments conducted under representative network protocol environments show that the in-kernel CNN inference framework reduces average inference latency by approximately 86.8% and lowers CPU usage by 18.7% without sacrificing accuracy. These results demonstrate the effectiveness and feasibility of the proposed approach for achieving efficient and low-overhead intelligent inference on resource-constrained edge devices.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

In-Kernel CNN Inference for Edge Devices: An eBPF-Based Approach to Low-Latency and Resource-Efficient Processing

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Yaodong Zheng,
  • Junxing Zhang

摘要

With the rapid proliferation of application scenarios such as the Internet of Things (IoT), industrial automation, smart homes, and vehicular networks, edge devices are facing increasing demands for low-latency and high-performance intelligent data processing. However, typical edge nodes—including home gateways, edge servers, and Customer Premises Equipment (CPE)—often suffer from limited computational resources, constrained memory capacity, and tight energy budgets. These constraints make it difficult for traditional user-space inference frameworks to achieve millisecond-level responsiveness under high-frequency data streams. To address this issue, this paper proposes a lightweight in-kernel Convolutional Neural Network (CNN) inference framework based on the extended Berkeley Packet Filter (eBPF), aiming to enable real-time traffic identification and classification on edge devices. By embedding the CNN model directly into the eBPF program of the Linux kernel, the proposed method performs packet parsing, feature extraction, and model inference entirely in kernel space, thereby eliminating the overhead of context switches and system calls common in user-space solutions. Experiments conducted under representative network protocol environments show that the in-kernel CNN inference framework reduces average inference latency by approximately 86.8% and lowers CPU usage by 18.7% without sacrificing accuracy. These results demonstrate the effectiveness and feasibility of the proposed approach for achieving efficient and low-overhead intelligent inference on resource-constrained edge devices.