Scholarlink

<p>Edge intelligence, combining artificial intelligence with edge computing, is a promising emerging technology. However, its efficient operation inevitably faces the challenge of handling massive data. To address this challenge, multi-exit models and partitioning models have been developed. Notably, existing research on multi-exit deep neural networks (DNN) inference partitioning models predominantly focuses on optimizing the model itself, with less attention to the task scheduling complexity in multi-user scenarios. In view of this, this paper delves into the issues of DNN inference partitioning and task scheduling within a multi-user environment, proposing a multi-user multi-exit DNN inference partitioning and task scheduling strategy. This strategy takes into account factors such as task urgency and deadlines, optimizing DNN exit and partition points while enabling rational task scheduling to maximize task completion rates and inference accuracy. We design a double greedy algorithm-based inference task scheduling mechanism that efficiently schedules inference tasks for multiple users on edge servers, achieving optimal utilization of system resources. Additionally, we introduce the multi-agent proximal policy optimization method to further optimize the performance of multi-exit and partition models. Simulation experiments, compared to other mainstream scheduling algorithms, demonstrate that the proposed algorithm effectively improves task completion rates.</p>

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

Multi-user multi-exit DNN inference partitioning and task scheduling strategy

  • 期刊论文

  • 发表日期：

    2025年6月11日

查看全文
  • Xianzhong Tian,
  • Yucheng Li,
  • Xuhua Mao,
  • Luoming Zhang,
  • Ziyi Lin

摘要

Edge intelligence, combining artificial intelligence with edge computing, is a promising emerging technology. However, its efficient operation inevitably faces the challenge of handling massive data. To address this challenge, multi-exit models and partitioning models have been developed. Notably, existing research on multi-exit deep neural networks (DNN) inference partitioning models predominantly focuses on optimizing the model itself, with less attention to the task scheduling complexity in multi-user scenarios. In view of this, this paper delves into the issues of DNN inference partitioning and task scheduling within a multi-user environment, proposing a multi-user multi-exit DNN inference partitioning and task scheduling strategy. This strategy takes into account factors such as task urgency and deadlines, optimizing DNN exit and partition points while enabling rational task scheduling to maximize task completion rates and inference accuracy. We design a double greedy algorithm-based inference task scheduling mechanism that efficiently schedules inference tasks for multiple users on edge servers, achieving optimal utilization of system resources. Additionally, we introduce the multi-agent proximal policy optimization method to further optimize the performance of multi-exit and partition models. Simulation experiments, compared to other mainstream scheduling algorithms, demonstrate that the proposed algorithm effectively improves task completion rates.