Abstract: In recent years, researchers pay more attention to the power consumption controlling in clusters. People have tried hard to reduce the power consumption of clusters. However, managing power is more important than reducing it. Power consumption is considered as an item which can be managed in clusters. For example, clients give the power consumption that they need and can afford. Managers use strategies to accomplish tasks submitted by clients and satisfy the power constraints given by clients. This paper designs two power controlling and predicting approaches based on HDFS (Hadoop distributed file system) clusters, sets power budget to every node in clusters and sets dynamic and non-uniform power constraints to control power consumption of the whole clusters. Beside this, this paper designs a stepwise multiple linear model to predict power and also tests and verifies these two approaches of predictive power controlling.

Key words: power predictive model, linear regression, predictive

摘要： 近几年来，研究人员越来越重视集群中的功率消耗控制问题。众多研究人员都致力于功率消耗的降低与节约，然而能耗管理比单纯的能耗节约显得更加重要。将功率看成是可以管理和调度的资源之一，用户做出一个能耗预算，而管理者则在用户的能耗预算下，帮助其用合理的调度算法和功率限制策略完成用户提交的任务。设计与实现了两种应用于HDFS（Hadoop distributed file system），有效控制与预测功率的算法和策略，结合给每个节点设置功率预算的方法，实现功率的非均衡的动态分配，以控制整个集群功率消耗，从而限制能耗。此外，还通过逐步线性回归得到的功率模型来优化功率管理策略，并对两种功率预测控制策略进行了分析比较

关键词: 能耗预测模型, 线性回归, 限制下预测控制, MapReduce