Scholarlink

<p>The rapid expansion of artificial intelligence (AI) and cloud services has significantly increased data center energy demand, motivating efficient scheduling strategies on DVFS-enabled heterogeneous platforms. Within the Advanced Configuration and Power Interface (ACPI) framework, <i>Dynamic Power Management</i> (DPM) and <i>Dynamic Voltage and Frequency Scaling</i> (DVFS) reduce static and dynamic energy consumption; however, applying them to dependent-task workflows under deadline constraints introduces complex trade-offs among static, frequency-independent, and frequency-dependent energy components. To address these challenges, this article presents <i>Scheduling for Energy Saving</i> (<i>S4eS</i>), a compile-time energy-aware scheduling algorithm with provably lower-bound time complexity. <i>S4eS</i> integrates a DPM module that selects compute nodes for switch-off using a theory-grounded criterion based on frequency-independent energy, and a DVFS module that determines task-level frequency scaling to maximize net energy savings while respecting deadlines. Theoretical analysis establishes the lower-bound complexity of both modules, ensuring scalability for large-scale environments. Experimental results on synthetic and real-world workflows show that <i>S4eS</i> achieves average energy savings of 20.08%, outperforming related methods while maintaining reduced computational overhead.</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

S4eS: a lower-bound complexity scheduling algorithm for energy savings on DVFS-enabled platforms

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月16日

查看全文
  • Tarek Hagras,
  • Gamal A. El-Sayed

摘要

The rapid expansion of artificial intelligence (AI) and cloud services has significantly increased data center energy demand, motivating efficient scheduling strategies on DVFS-enabled heterogeneous platforms. Within the Advanced Configuration and Power Interface (ACPI) framework, Dynamic Power Management (DPM) and Dynamic Voltage and Frequency Scaling (DVFS) reduce static and dynamic energy consumption; however, applying them to dependent-task workflows under deadline constraints introduces complex trade-offs among static, frequency-independent, and frequency-dependent energy components. To address these challenges, this article presents Scheduling for Energy Saving (S4eS), a compile-time energy-aware scheduling algorithm with provably lower-bound time complexity. S4eS integrates a DPM module that selects compute nodes for switch-off using a theory-grounded criterion based on frequency-independent energy, and a DVFS module that determines task-level frequency scaling to maximize net energy savings while respecting deadlines. Theoretical analysis establishes the lower-bound complexity of both modules, ensuring scalability for large-scale environments. Experimental results on synthetic and real-world workflows show that S4eS achieves average energy savings of 20.08%, outperforming related methods while maintaining reduced computational overhead.