Energy-Efficient Thermal Control for 2U Air-Cooled Servers via MADDPG and Online Learning
Abstract
This research proposes an intelligent thermal control methodology that integrates the Multi-agent Deep
Deterministic Policy Gradient (MADDPG) algorithm with an Online Learnin. The primary objective is to
address the critical challenges of thermal coupling and excessive energy consumption in high-power servers
operating under dynamic workloads. In this study, a 2U air-cooled server platform—supporting a maximum
power input of 2200 W and equipped with six cooling fans (each featuring a maximum airflow of 60 CFM)—
serves as the experimental backbone.
The proposed control architecture treats each cooling fan as an autonomous yet collaborative agent, enabling the
system to effectively manage the complex Multi-Input Multi-Output (MIMO) thermal characteristics inherent in
multi-heat-source environments. Experimental results demonstrate that the standalone MADDPG algorithm
achieves an 8.33% reduction in fan energy consumption compared to traditional Proportional-Integral (PI)
control. Furthermore, by incorporating the Online Learning mechanism, the control strategy can continuously
adapt and update its parameters in response to real-time system states. This integration further optimizes
performance, leading to a total energy saving of 11.11% relative to conventional PI control. The findings
underscore the potential of reinforcement learning-based approaches in enhancing the energy efficiency and
thermal stability of next-generation data center hardware.
Deterministic Policy Gradient (MADDPG) algorithm with an Online Learnin. The primary objective is to
address the critical challenges of thermal coupling and excessive energy consumption in high-power servers
operating under dynamic workloads. In this study, a 2U air-cooled server platform—supporting a maximum
power input of 2200 W and equipped with six cooling fans (each featuring a maximum airflow of 60 CFM)—
serves as the experimental backbone.
The proposed control architecture treats each cooling fan as an autonomous yet collaborative agent, enabling the
system to effectively manage the complex Multi-Input Multi-Output (MIMO) thermal characteristics inherent in
multi-heat-source environments. Experimental results demonstrate that the standalone MADDPG algorithm
achieves an 8.33% reduction in fan energy consumption compared to traditional Proportional-Integral (PI)
control. Furthermore, by incorporating the Online Learning mechanism, the control strategy can continuously
adapt and update its parameters in response to real-time system states. This integration further optimizes
performance, leading to a total energy saving of 11.11% relative to conventional PI control. The findings
underscore the potential of reinforcement learning-based approaches in enhancing the energy efficiency and
thermal stability of next-generation data center hardware.
