Scholarlink

<p>Edge-based IoT systems must execute delay-sensitive tasks under dynamic workloads and limited computational resources, where reactive scheduling often leads to high deadline miss rates. Most existing reinforcement learning based edge schedulers rely primarily on instantaneous system states, limiting their ability to anticipate short-term congestion. This paper studies joint task migration and CPU allocation in IoT edge computing under deadline constraints. The problem is formulated as a Markov Decision Process, and a lightweight Transformer encoder is used to summarize short sequences of recent system states, capturing near-term workload and queue dynamics. The encoded temporal representation is integrated into an Asynchronous Advantage Actor-Critic (A3C) framework, enabling the agent to make scheduling decisions with predictive context. The learned policy jointly selects task migration targets and CPU allocation levels using a reward function that balances deadline satisfaction, latency, energy consumption, and migration overhead. Simulation results under bursty workloads show that the proposed method reduces deadline miss rates by up to 35%, lowers P99 latency by approximately 40%, and improves energy efficiency compared to standard A3C, PPO, and heuristic schedulers. These results indicate that augmenting DRL-based edge schedulers with Transformer-based temporal state encoding improves deadline reliability under dynamic IoT workloads. The source code is available at: <a href="https://github.com/Kiruthika-1919/transformer-a3c-edge-scheduler.git.">https://github.com/Kiruthika-1919/transformer-a3c-edge-scheduler.git.</a></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

Deadline-aware task migration and resource allocation in IoT edge computing via transformer-augmented A3C

  • 期刊论文

  • 发表日期：

    2026年5月8日

查看全文
  • Kiruthika Rayar Angammal,
  • Qaisar Ali,
  • Muhammad Umar Chaudhry,
  • Ihsan Ullah

摘要

Edge-based IoT systems must execute delay-sensitive tasks under dynamic workloads and limited computational resources, where reactive scheduling often leads to high deadline miss rates. Most existing reinforcement learning based edge schedulers rely primarily on instantaneous system states, limiting their ability to anticipate short-term congestion. This paper studies joint task migration and CPU allocation in IoT edge computing under deadline constraints. The problem is formulated as a Markov Decision Process, and a lightweight Transformer encoder is used to summarize short sequences of recent system states, capturing near-term workload and queue dynamics. The encoded temporal representation is integrated into an Asynchronous Advantage Actor-Critic (A3C) framework, enabling the agent to make scheduling decisions with predictive context. The learned policy jointly selects task migration targets and CPU allocation levels using a reward function that balances deadline satisfaction, latency, energy consumption, and migration overhead. Simulation results under bursty workloads show that the proposed method reduces deadline miss rates by up to 35%, lowers P99 latency by approximately 40%, and improves energy efficiency compared to standard A3C, PPO, and heuristic schedulers. These results indicate that augmenting DRL-based edge schedulers with Transformer-based temporal state encoding improves deadline reliability under dynamic IoT workloads. The source code is available at: https://github.com/Kiruthika-1919/transformer-a3c-edge-scheduler.git.