摘要
为了准确描述I/O行为,以缓解复杂应用环境下I/O突发聚集所引起的海量存储系统性能的瓶颈,着眼于时间序列范畴,从trace样本数据的获取、I/O负载模型的选择、I/O负载模型的参数估计和负载模型的拟和程度的评估等4个方面,系统地阐述了一种基于trace时间序列的负载模型构造方法,并通过实际应用对其进行检验.针对一类科学计算应用trace的样本数据,分布检验显示其具有高斯特征,相关性研究表明分形布朗运动模型适合相应trace样本数据的合成.对高斯的正态模型和分形布朗运动模型分别进行参数估计,合成负载和实际数据之间的单因素方差分析结果表明:分形布朗运动模型能更有效地描述存储负载中的I/O突发行为.
In order to accurately characterize I/O behaviors and remove the bottleneck problem in mas sive storage system caused by I/O bursty on complex applications, focused on a time dependence perspective, an approach of structuring synthetic model based on time series, i. e. , collecting sample trace, selecting I/O workload models, estimating the parameters of workload models, and evaluating the fitting degree, was proposed and implemented. Then it was examined by real scientific computing trace. Distribution examination indicates it exhibits the Gaussian property, and correlation study shows fractal brown motion (FBM) model is appropriate for the synthesis of real trace. The parameters of the Gaussian norm and FBM models were estimated, respectively. The experiment results of the analysis of variance between real and synthetic data show that FBM can more effectively capture the anomaly behaviors in I/O workloads.
出处
《华中科技大学学报(自然科学版)》
EI
CAS
CSCD
北大核心
2013年第3期48-51,共4页
Journal of Huazhong University of Science and Technology(Natural Science Edition)
基金
国家科技支撑计划资助项目(2012BAD35B08)
中央高校基本科研业务费专项资金资助项目(XDJK2012A006)
西南大学博士基金资助项目(SWU111015)
关键词
负载模型
时间序列
自相似
复杂应用
自相关
workload model
time series
self-similar
complex applications
auto-correlation