Acceleration of Analysis Processing on Decentralized Performance Profiling System Using Virtual Machines

To detect the performance anomaly of a computer, as a structure for continuous performance profiling, decentralization of the performance profiling system using virtual machines has been proposed. Moreover, there have already been evaluation results reported regarding overhead, including data storing, and data sampling stall time. On the other hand, for continuous performance profiling, the continuous processing of performance profiling is needed, including not only data sampling and data storing but also analysis processing. Therefore, first, this paper describes a relationship condition among data sampling time, data storing time, and analysis processing time as the necessary condition for continuous performance profiling on a decentralized performance profiling system. Second, in order to satisfy the relationship condition, we propose a concurrent operation technique as the acceleration method of analysis processing for a decentralized performance profiling system. Finally, this paper presents quantitative evaluations of the proposed method, including the case of a multi-VMM environment.