Performance Evaluation In Multiple Virtual Machines Based On The Virtualization Of Hardware Performance Counters

With the wide application of virtualization technology, the demand for performance analysis and system diagnosis in virtualization environment is increasing. For example, users in different virtual machines may need to obtain the performance issues of programs in respective virtual machine. A performance analysis approach is based on the virtualization of hardware performance counters, and it obtains performance information of programs during its execution by monitoring respective CPU events. However, hardware performance counters are a kind of special purpose registers, and the count of them is limited. For example, Xenoprof, the toolkit in Xen virtual machine environment, can make multiple virtual machines monitor the same CPU events in a session simultaneously, but monitoring different CPU events in different virtual machines simultaneously is not allowed. Therefore, it can't satisfy the individual profiling need of multiple users in Xen virtual machine environment. Considering the issues above, this paper puts forward a kind of research about the performance evaluation approach for multiple virtual machines based on the virtualization of hardware performance counters.The virtualization of hardware performance counters can make multiple virtual machines monitor different CPU events simultaneously. To ensure the profiling precision, we put read/write operations of hardware performance counters into the time slices of Virtual Machine Monitor (VMM). Besides, we use the sampling mechanism to obtain sampling data in kernel level, and then distribute data into respective buffers of virtual machines, and eventually user-space daemons in each virtual machine are in charge of the analysis of sampling data to get detailed performance information. So, users in different virtual machines can monitor different CPU events simultaneously without feeling the existence of other users, which provides great convenience for programming and performance evaluation in virtual machine environment.From the test results, we can see that the system can well satisfy the demands on the function, and it can make multiple virtual machines monitor different CPU events simultaneously without feeling the existence of other users; Meanwhile, it also perform well on the performance, and the test results generated by the system are close to what generated by Xenoprof under the same conditions.