Scholarlink

<p>There is an immediate necessity of innovative and adaptable strategies to optimize key performance metrics such as latency, throughput, energy efficiency and resource utilization since job scheduling is becoming more complex in the Edge IoT networks. There are problems with conventional scheduling methods, which are unable to meet the real-time processing demands or dynamically assigning jobs, e.g., Reinforcement Learning-based Scheduling (RL-S) and Genetic Algorithm-based Scheduling (GA-S). To address these constraints we present Deep Learning based Task Scheduling (DL-TS) algorithm which is incorporated on the Convolutional Neural Network (CNN) to optimally allocate computing jobs to edge nodes. Training of the CNN model is performed on a simulation-generated dataset of 100,000 IoT task records, produced by 1000 heterogeneous IoT devices distributed across 10 edge nodes and 5 cloud servers, covering computational, communication-heavy, and mixed workloads under variable Poisson arrival rates. To ensure that the model is ready to attend to the changeable nature of the Edge IoT environments, essential scheduling parameters such as CPU load, task size, bandwidth, priority, and response time have been part of the dataset. In our test results, the proposed DL-TS approach performs significantly better when compared to standard methods. Specifically, it reduces network latency by 34.8% and job completion time by 27.4% as compared to GA-S and RL-S. The framework also attains resource utilization more effectively and with less energy consumption as it promotes throughput 22.1%. Statistical validation across 30 independent runs using paired Student’s t-tests with Holm–Bonferroni correction confirms that all reported improvements over GA-S, RL-S, MLP-S, LSTM-S, and Transformer-S baselines are significant at <i>p</i> &lt; 0.01 with large effect sizes (Cohen’s d &gt; 0.8). A scalability study further shows that DL-TS sustains its performance up to 50 edge nodes and 10,000 IoT devices, while maintaining per-task inference latency below 2.0 ms on resource-constrained edge hardware. The outcomes consistently prove the high efficiency of the proposed approach, which means that it can be applied in the Edge IoT networks in practice. The proposed DL-TS approach suitable for adapting distributed environments and provides a scalable and effective real-time task scheduling model.</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

Deep Learning-Based Task Scheduling in Edge IoT Networks: A CNN-Driven Approach for Optimized Resource Management and Performance Enhancement

  • 期刊论文

  • 发表日期：

    2026年5月20日

查看全文
  • Himabindu Sathyaveti,
  • C. Gomathy

摘要

There is an immediate necessity of innovative and adaptable strategies to optimize key performance metrics such as latency, throughput, energy efficiency and resource utilization since job scheduling is becoming more complex in the Edge IoT networks. There are problems with conventional scheduling methods, which are unable to meet the real-time processing demands or dynamically assigning jobs, e.g., Reinforcement Learning-based Scheduling (RL-S) and Genetic Algorithm-based Scheduling (GA-S). To address these constraints we present Deep Learning based Task Scheduling (DL-TS) algorithm which is incorporated on the Convolutional Neural Network (CNN) to optimally allocate computing jobs to edge nodes. Training of the CNN model is performed on a simulation-generated dataset of 100,000 IoT task records, produced by 1000 heterogeneous IoT devices distributed across 10 edge nodes and 5 cloud servers, covering computational, communication-heavy, and mixed workloads under variable Poisson arrival rates. To ensure that the model is ready to attend to the changeable nature of the Edge IoT environments, essential scheduling parameters such as CPU load, task size, bandwidth, priority, and response time have been part of the dataset. In our test results, the proposed DL-TS approach performs significantly better when compared to standard methods. Specifically, it reduces network latency by 34.8% and job completion time by 27.4% as compared to GA-S and RL-S. The framework also attains resource utilization more effectively and with less energy consumption as it promotes throughput 22.1%. Statistical validation across 30 independent runs using paired Student’s t-tests with Holm–Bonferroni correction confirms that all reported improvements over GA-S, RL-S, MLP-S, LSTM-S, and Transformer-S baselines are significant at p < 0.01 with large effect sizes (Cohen’s d > 0.8). A scalability study further shows that DL-TS sustains its performance up to 50 edge nodes and 10,000 IoT devices, while maintaining per-task inference latency below 2.0 ms on resource-constrained edge hardware. The outcomes consistently prove the high efficiency of the proposed approach, which means that it can be applied in the Edge IoT networks in practice. The proposed DL-TS approach suitable for adapting distributed environments and provides a scalable and effective real-time task scheduling model.