面向高能物理分级存储的文件访问热度预测  

File Access Popularity Prediction for Hierarchical Storage for High-Energy Physics

作　　者：程振京 汪璐[1,2] 程耀东 陈刚[1] 胡庆宝 李海波[1,2] CHENG Zhenjing;WANG Lu;CHENG Yaodong;CHEN Gang;HU Qingbao;LI Haibo(Institute of High Energy Physics,Chinese Academy of Sciences,Beijing 100049,China;University of Chinese Academy of Sciences,Beijing 100049,China;Tianfu Cosmic Ray Research Center,Institute of High Energy Physics,Chinese Academy of Sciences,Chengdu 610041,China)

机构地区：[1]中国科学院高能物理研究所,北京100049 [2]中国科学院大学,北京100049 [3]中国科学院高能物理研究所天府宇宙线研究中心,成都610041

出　　处：《计算机工程》2021年第2期126-132,共7页Computer Engineering

基　　金：国家重点研发计划(2017YFB0203200);国家自然科学基金(11675201,11805226,11805223)。

摘　　要：高能物理计算是典型的数据密集型计算,其主要采用基于文件的分级存储方案,根据访问热度的不同将数据存储于不同性能的存储设备上,然而当前数据热度预测采用基于人工经验的启发式算法,准确率较低。提出一种借助长短期记忆网络预测文件未来访问热度的方法,包括网络结构设计、训练和预测算法等。该方法通过划分动态时间窗口构造文件访问特征的时序序列,预测不同数据的访问趋势。在LHAASO高能物理实验数据集上的实验结果表明,与SVM、MLP等算法相比,该方法预测准确率提升了30%左右,具有更强的适用性。Computing for high-energy physics is typically data-intensive. It mainly adopts file-based hierarchical storage solutions where data is allocated based on the access popularity to storage devices with different performances.The existing schemes of data popularity prediction generally adopt a heuristic algorithm based on artificial experience,whose prediction accuracy is low.This paper proposes a method of predicting future access popularity using Long ShortTerm Memory(LSTM) network,which consists of network structure design,training,and prediction algorithms. The method divides the dynamic time window to construct a time series of file access features,and on this basis predicts the access trends of different data.Experimental results on the data set of LHAASO high-energy physics experiments show that compared with SVM,MLP and other algorithms,the proposed method increases the prediction accuracy by about30%,and it has stronger applicability.

关 键 词：分级存储 文件访问特征 时序数据 长短期记忆网络 文件访问热度 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]