机群系统中空闲结点的功耗管理

Power Management of Idle Nodes in Clusters

针对机群系统中存在的大量空闲活跃结点所造成的严重能耗浪费,提出空闲结点的cache式动态功耗管理模型,即利用结点多级休眠机制,将空闲结点划分为不同休眠等级的结点集合,每级休眠状态对应一级结点储备cache,力求获得近似活跃状态的系统响应速率,以及近似最深休眠状态的能耗节省。基于cache式功耗管理模型,综合能耗与响应速率两个因素,设计了空闲结点在不同休眠状态之间的动态升降级算法、基于储备池的资源结点分配与回收算法以及储备额阈值自适应算法,以在保证系统响应速率的同时降低系统能耗。实验表明,提出的空闲结点cache式功耗管理技术在作业相对延迟仅增加0.99%的代价下,系统空闲结点功耗降低69.51%,优化效果显著。

Existence of massive active idle nodes causes huge energy waste in large scale systems.Cache-style power management for idle nodes was proposed to schedule the power states of idle nodes.According to their different sleep states,idle nodes are placed into multiple groups with corresponding sleep states.It is expected to achieve a system response speed similar to the active state and a power saving similar to the deepest sleep state.The idle nodes are dynamically transformed between different sleep groups.Assuring response speed of system,idle node is put into a sleep state as deep as possible.In our experiments,CPMI conserves the power consumption of idle nodes by 69.51% with the cost of relative slowdown only by 0.99%.