Abstract: This paper proposes a multi-level cluster power management，reducing power consumption of the cluster system with less effect on performance.This management can be divided into power management of cluster level and local.Cluster level mechanism presents an ondemand-start strategy based on self-learning load forecasting algorithm，providing computing resources according to the load.Local mechanism suggests an enhanced-conservative governor algorithm to improve the sensitivity of the frequency adjustment when load drops.The experiments show that this multi-level power management is more effective than other strategies for reducing overall system power consumption.

Key words: multi-level cluster power management, self-learning load forecasting algorithm, ondemand-start, frequency adjustment

摘要： 多层次的集群功耗管理方法，在不明显影响系统性能的前提下，降低集群系统的功耗。该管理方法分为集群层次的功耗管理和本地节点层次的功耗管理。集群层次的功耗管理基于自学习负载预测的按需启动策略，根据作业的负载提供计算资源。本地节点层次的功耗管理针对负载下降产生的节点空闲问题，提出了Enhanced-conservative调控器算法，提高了负载下降时频率调整的敏感度。测试实验数据表明，该策略比其他策略能更有效的降低整个系统的功耗。

关键词: 多层次功耗管理, 自学习负载预测, 按需启动, 频率调整