一种分段集群异常作业预测方法  

A prediction method of staged cluster anomaly job

作　　者：谢丽霞[1] 汪子荧 XIE Lixia;WANG Ziying(School of Computer Science and Technology,Civil Aviation University of China,Tianjin 300300,China)

机构地区：[1]中国民航大学计算机科学与技术学院

出　　处：《大连理工大学学报》2019年第4期427-433,共7页Journal of Dalian University of Technology

基　　金：国家自然科学基金民航联合研究基金资助项目(U1833107);国家科技重大专项资助项目(2012ZX03002002);中央高校基本科研业务费专项资金资助项目(ZYGX2018028)

摘　　要：针对现有集群异常作业预测方法预测效率低、预测时间长的问题,提出一种分段集群异常作业预测(SCAJP)方法.该方法分为离线预测和在线预测两个阶段:离线预测阶段,依据作业子任务的静态特征对子任务终止状态进行预测,并仅在线预测此阶段的正常子任务所属作业;在线预测阶段,在计算作业子任务动态特征的同时,采用改进门控递归单元(IGRU)神经网络根据动态特征实时预测任务终止状态是否异常.两个阶段的最后均根据作业与其子任务的相关性检索异常作业,实现对异常作业的预测.实验结果表明,该方法在灵敏度、精确度和预测时间方面明显优于其他方法.Aiming at the problems of low prediction efficiency and long prediction time of the existing cluster anomaly job prediction methods, a staged cluster anomaly job prediction (SCAJP) method is proposed. This method is divided into offline stage and online stage. The final state of the job′s sub-tasks is predicted according to their static features in the offline stage, then the prediction is only done for the job to which the normal sub-task belongs. In online stage, while calculating the dynamic features of the job′s sub-tasks, the improved gated recurrent unit (IGRU) neural network is used to predict whether the task termination status is anomaly according to the dynamic features in real time. At the end of the both stages, the anomaly job is obtained based on the relevance between the job and its sub-tasks to finish the prediction of the anomaly job. The experimental results show that this method outperforms other methods in terms of sensitivity, accuracy and prediction time obviously.

关 键 词：集群异常作业 分段预测 实时预测 动态特征 门控递归单元 

分 类 号：TP309.7[自动化与计算机技术—计算机系统结构]