基于性能分析的自适应插桩框架

AN ADAPTIVE INSTRUMENTATION FRAMEWORK BASED ON PERFORMANCE ANALYSISWang Zipeng1,2, Ren Zhongshan1,3, Zhang Shudong1,2, Hu Jianya1,2

摘要:

插桩技术用于跟踪获取软件系统的运行时信息,是软件性能管理工具中不可或缺的一个部分。目前,存在的各类插桩工具所产生的插桩点在目标程序运行的过程中往往是不可改变的。在实际的程序异常、错误检测中,用户关心的程序代码的位置在不同阶段往往不同,因此目标程序运行的不同阶段所需要获取信息的插桩点位置是不同的,如果插桩点过多会导致插桩工具浪费系统资源,产生系统资源消耗过大的问题；插桩点过少则会导致无法准确定位目标软件发生异常位置。针对以上问题,使用基于线性回归和K-Means的分析模型,分析目标软件的性能数据,在其运行过程中动态地改变插桩点,尽可能的减少资源的消耗；另外,采用朴素贝叶斯分类模型对插桩类进行筛选,减少植入插桩点的类,可降低插桩带来的资源消耗。实验表明：与传统的工具相比,使用此插桩框架进行监控,被测网页的平均响应时间减少6.88%,同时对目标程序的干扰更小。

Abstract:

Instrumentation is an indispensable part of the application performance management (APM) and used for tracking and obtaining the information of software system. Nowadays, there are various instrumentation tools, and the probes of the tools are unchanged during target application operation. However, in daily exception and bug detection, users require different code monitoring in different periods of a target application operation, and too many probes may have the result that the instrumentation tool wastes and consumes more system resources; too few probes may result that APM can’t locate accurate positioning failure of target software. We proposes a new adaptation instrumentation framework based on analyzing performance data and carries out adaption instrumentation based on liner regression and K-Means, changes the probes dynamically during operation and reduces the resource waste as much as possible. What’s more, this paper uses Na?ve Bayes to filter Java classes for reducing instrumentation classes and overhead; In comparison to the traditional tools, this paper leads to low overhead and disturbance through the experiment, average page response time reduced by about 6.88%.