面向按需供给的资源需求滤波估算方法

Filter Based Resource Demand Estimation for On-demand Provision

随着按需供给资源使用模式的推广,软件的资源需求已成为资源优化控制的重要属性.监测和估算是目前常用的资源消耗获取方法,但监测工具难以在运行时准确度量短任务的资源需求,回归分析方法又因受到多元共线性和不确定性因素的影响,导致其取值精度下降.本文提出了一种基于Kalman滤波的资源需求估算方法.该方法建立了可度量属性集与不可度量的资源需求间的关联,并利用滤波过滤度量过程中的噪声,达到降低估算误差的目的.基准测试的结果表明,通过合理的设置滤波参数,本方法能够快速逼近真实值,且平均误差小于8%.

As the development of demand resource provision, resource demands of software is becoming one of the most important attributes of resource management. Measurement and estimation are widely used in fetching the demands. However, it is hard to measure the short jobPs resource demands by current measurement tools, and the regression methods suffer from the well-studied problem of multicollinearity. Therefore, the estimated results are not confident. In order to improve the estimation precision, we propose a Kalman filter based approach, which can predict the unobservable attribute by observable attributes, and filter the noise existing in the measurement. At last, we test our approach with a benchmark and compare the relative errors, which can demonstrate that with the reasonable parameters, our approach can get close to the real demands quickly, and get the estimated value with the mean error less than 8 %.