Scholarlink

<p>Cloud servers, as computing units in data centers, have become a focal point for energy conservation due to their enormous energy consumption. With the adaptive decision-making capability, reinforcement learning (RL) has demonstrated great potential in the energy optimization of cloud servers. Some studies have proposed RL-based methods for dynamic voltage and frequency scaling (DVFS) and dynamic fan speed management to improve server energy efficiency. However, the vast exploration space and the sparsity of rewards necessitate substantial training time to learn the RL policy. Hence, this paper proposes an adversarial RL-based method, named ARDF, to optimize the RL training in realistic server environments. In ARDF, a differential action rewarding approach based on baseline comparison is designed to address the multi-step, short-sighted decision problem encountered in server DVFS. Additionally, the expert experience is leveraged to guide the agent’s learning through an adversarial approach, thereby expediting the acquisition of a high-quality policy. The experiments utilize benchmark tests to validate the effectiveness of ARDF on the Huawei Taishan 200 server. Results show ARDF improves energy efficiency by 7.3% over the server’s built-in energy-saving mode and by 0.9% over the best baseline, while maintaining robust adaptability under varying thermal constraints.</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

An adversarial reinforcement learning method to enhance server energy efficiency via DVFS and dynamic fan speed management

  • 期刊论文

  • 发表日期：

    2025年6月28日

查看全文
  • Wenjun Lin,
  • Jianpeng Lin,
  • Jiachun Li,
  • Mian Guo,
  • Ligang He,
  • Weiwei Lin

摘要

Cloud servers, as computing units in data centers, have become a focal point for energy conservation due to their enormous energy consumption. With the adaptive decision-making capability, reinforcement learning (RL) has demonstrated great potential in the energy optimization of cloud servers. Some studies have proposed RL-based methods for dynamic voltage and frequency scaling (DVFS) and dynamic fan speed management to improve server energy efficiency. However, the vast exploration space and the sparsity of rewards necessitate substantial training time to learn the RL policy. Hence, this paper proposes an adversarial RL-based method, named ARDF, to optimize the RL training in realistic server environments. In ARDF, a differential action rewarding approach based on baseline comparison is designed to address the multi-step, short-sighted decision problem encountered in server DVFS. Additionally, the expert experience is leveraged to guide the agent’s learning through an adversarial approach, thereby expediting the acquisition of a high-quality policy. The experiments utilize benchmark tests to validate the effectiveness of ARDF on the Huawei Taishan 200 server. Results show ARDF improves energy efficiency by 7.3% over the server’s built-in energy-saving mode and by 0.9% over the best baseline, while maintaining robust adaptability under varying thermal constraints.