An Attention-Driven Deep Reinforcement Learning Framework for Energy-Efficient and Service-Level Agreement-Aware Cloud Task Scheduling | ||
| AUT Journal of Electrical Engineering | ||
| مقالات آماده انتشار، پذیرفته شده، انتشار آنلاین از 01 تیر 1405 اصل مقاله (2.3 M) | ||
| نوع مقاله: Research Article | ||
| شناسه دیجیتال (DOI): 10.22060/eej.2026.25455.5936 | ||
| نویسندگان | ||
| Rahul Bhatt1؛ Ritika Mehra2؛ Kamal Upreti* 3 | ||
| 1School of Engineering & Computing, Dev Bhoomi Uttarakhand University, Dehradun, Uttarakhand, India | ||
| 2School of Engineering & Computing, Dev Bhoomi Uttarakhand University Dehradun, Uttarakhand, India | ||
| 3Department of Computer Science, Christ University, Delhi NCR Campus, Ghaziabad, India | ||
| چکیده | ||
| Dynamic cloud and edge-cloud platforms require task schedulers that can respond to stochastic workloads while minimizing energy use and preserving service-level agreement reliability. This study proposes an attention-driven deep reinforcement learning framework for energy-efficient and service-level agreement-aware cloud task scheduling. The framework combines lightweight convolutional neural network and long short-term memory-based spatial-temporal feature extraction with a multi-head self-attention actor-critic decision module. The convolutional neural network and long short-term memory components capture local virtual machine workload patterns, whereas self-attention models global virtual-machine-to-virtual-machine dependencies for parallel and context-aware scheduling. The scheduling problem is formulated as a Markov decision process using a 242-dimensional virtual-machine-level state representation, probabilistic virtual-machine-to-host assignment actions, and a multi-objective reward function covering makespan, energy consumption, operational resource cost, and service-level agreement penalties. Experiments were conducted in a heterogeneous CloudSim environment with 100 hosts and 100 virtual machines. The proposed framework achieved a normalized makespan of approximately 0.85, a 14.4% reduction in total energy consumption, consistently low service-level agreement violation behavior, and controlled migration activity. Logged analysis further showed a response time of 10.0000 milliseconds per completed task or virtual machine event, supporting interval-based real-time feasibility. Cost is treated as an operational reward component, not as a standalone billing analysis. | ||
| کلیدواژهها | ||
| Cloud Computing؛ Task Scheduling؛ Deep Reinforcement Learning؛ Multi-Head Self-Attention؛ Actor-Critic Learning؛ Energy Efficiency؛ Service-Level Agreement-Aware Scheduling؛ CloudSim | ||
|
آمار تعداد مشاهده مقاله: 101 تعداد دریافت فایل اصل مقاله: 206 |
||