数据库服务器系统软功率计

Constructing a software power-meter for database server

功率建模是功率感知DBMS的基础.针对数据库服务器系统整机系统进行功率建模,构建其软功率计.模型利用处理器、磁盘(包括硬盘和固态盘)和内存等主要部件的活动信息,包括其利用率与核频率,其余部件产生功率作为背景功率,来构建模型.为提高模型的准确性,对各部件、可用参数对模型准确性进行全面分析.实验表明不同的模型准确性、鲁棒性存在一定的差异,大部分模型的相对误差在5%~8%之间,个别没有利用足够信息的模型相对误差在12%左右,利用全部信息的模型具有较好准确性和稳定性.

Power modeling is fundamental to power aware DBMS.This paper aims to model the full-system power for a DBMS server and construct a software power meter for such systems.Our power models exploit the activity information of the main components（CPU,disk and memory）,including their utilizations and CPU core frequency.The power generated by remaining components is treated as the background power.To improve the accuracy of power models,we investigate various combinations of components and power factors in models.Our extensive experiments have revealed that different models have different accuracy and robustness.Most of the models achieve the relative error rate of 5% to 8% while some models,which failed to take advantage of available information,have larger relative error rate about 12%.The model that uses all information has better accuracy and stability.