机群系统基于预测的软件Rejuvenation策略研究  

Analysis of Prediction-based Software Rejuvenation in Cluster Systems

作　　者：范新媛[1] 徐国治[1] 蒋乐天[1] 

机构地区：[1]上海交通大学电子工程系,上海200030

出　　处：《计算机仿真》2004年第11期250-253,共4页Computer Simulation

摘　　要：处于运行阶段的软件往往存在老化现象 ,即软件的性能随着时间的推移逐渐下降 ,甚至导致突发的系统停机。为了抵消软件老化带来的消极影响 ,人们提出了一种称为软件rejuvenation的预防性容错策略。该策略通过主动的停止程序运行、清理程序的内部环境等步骤使重启后的软件回到正常的初始状态 ,避免了老化引起的突发性失效 ,因而可以提高系统的可用性。现有的软件rejuvenation包括定时和预测两种方法。该文把基于预测的rejuvenation策略引入机群 ,并综合考虑了机群系统前端的分配器和后端的服务器 ,对这两类实体分别实施rejuvenation ,以改善系统性能。该文构建了在rejuvenation策略下整个机群的SRN模型 ,以分析不同的预测概率对系统性能的影响。仿真结果表明该策略可以有效的提高整个机群系统的可用性。Recent study shows that the phenomenon of software aging exists in many long-running software systems, which will cause the performance degradation and even crash/hang failure of the systems. In order to counteract the negative effect of software aging, a preventive fault-tolerance technique named software rejuvenation has been proposed recently. It involves proactively terminating an application, cleaning its internal state and then restarting it to the refreshed state. Through these procedures, the unscheduled system failure due to aging can be effectively prevented.Software rejuvenation can be divided into time-based and prediction-based approaches. This paper introduces a prediction-based rejuvenation strategy into cluster systems and takes into account both the dispatcher and the worker pool in clusters. Rejuvenation is executed either on the dispatcher or a worker as soon as the failure probable state of the corresponding node is successfully detected. Based on this approach, a SRN model of the whole cluster is constructed to analyze the effect of different prediction coverage value to system availability. Numerical results show that this rejuvenation strategy can significantly increase the availability of the whole cluster system.

关 键 词：软件老化 机群 分配器/服务器架构 随机回报网 

分 类 号：TP206.3[自动化与计算机技术—检测技术与自动化装置]