基于SVM分类的云集群失败作业主动预测方法  

Predicting Job Failure in Cloud Cluster: Based on SVM Classification

作　　者：刘春红[1,2] 韩晶晶[1] 商彦磊[2] LIU Chun-hong HAN Jin-jin SHANG Yan-lei(College of Computer and Information Engineering, Henan Normal University, Henan Xinxiang 453002, China State Key Laboratory of Networking and Switching Technology, Beijing University of Posts and Telecommunications, Beijing 100876, China)

机构地区：[1]河南师范大学计算机与信息工程学院,河南新乡453002 [2]北京邮电大学网络与交换技术国家重点实验室,北京100876

出　　处：《北京邮电大学学报》2016年第5期104-109,共6页Journal of Beijing University of Posts and Telecommunications

基　　金：国家重点基础研究发展计划(973计划)项目(2012CB315802);国家关键技术研究与发展计划项目(2012BAH94F02);河南省科技厅基础与前沿技术研究项目(132300410430)

摘　　要：提出了一种使用支持向量机(SVM)模型预测作业终止状态的方法.以Google数据集为研究对象,首先分析作业终止状态的影响因素,提出使用作业的静态特征和动态特征作为终止状态分类的特征向量,选择SVM模型主动预测终止状态;然后从特征向量和分类模型2个层面对准确率、假负率、精确度指标进行验证.特征向量实验结果表明,基于静态和动态特征的SVM预测模型比单独使用静态特征和动态特征,分别提高0.94%、-0.01%、1.35%和9.08%、-1.36%、10.91%.分类模型的比较结果显示,SVM分类预测方法比传统的神经网络模型、朴素贝叶斯模型、逻辑回归模型的预测效果好.A job failure predicting method based on support vector machine（ SVM） model was presented. Google cluster traces were studied. The relevant factors of jobs failure were analyzed and the combination of the static and dynamic characteristic was chosen as the feature vectors. The SVM algorithm was chosen to predict termination status of the jobs. Experiments were conducted to compare different kinds of feature vectors and classification models with Google traces dataset in terms of the accuracy rate,false negative rate and precision rate. It is shown that the combination of static and dynamic features are0. 94%,-0. 01% and 1. 35% higher than the static features,and 9. 08%,-1. 36% and 8. 91%higher than the dynamic features. Experiments also demonstrate that the SVM model is superior to the traditional neural network extreme machine learning,naive Bayes and logistic regression model in these indexes.

关 键 词：失败作业预测 支持向量机模型 Google集群数据 

分 类 号：TN929.53[电子电信—通信与信息系统]