Reza NasiriGerdeh, Negin Hosseini, Rohollah Ehsani, Keyvan RahimiZadeh, Morteza AnaLoui
08/29/2017 11:09 PM
Computer Science
Server virtualization comforts deployment of Internet services and enables cloud service providers to improve resource utilization, fault tolerance, and energy efficiency by consolidating virtual servers within a shared environment. Although deploying services in virtualized environment brings benefits to service providers in terms of service agility, functionality, and reusability in comparison to traditional architecture, overhead of mediating hypervisor and contention amongst hosted virtual machines to possess virtualized resources can adversely affect the performance of applications in virtualized environments. In this paper, we provide a detailed profiling-based performance investigation of Web application in Xen virtualized environment to recognize the major sources of performance degradation of Web application in Xen. Our experimental results indicate that the main cause of the performance degradation of a single Web application in Xen is higher CPU utilization not higher disk utilization of the system; detailed profiling of the performance events occurred at the system-level identifies the most CPU consuming and the highest L3 cache miss functions of Xen and Dom0 kernels that contribute the most to this performance degradation. Additionally, the experiments show that by increasing the number of cores in Xen environment, the number of L3 cache misses occurred at the system decreases significantly which leads to response time improvement of the Web application in Xen. Moreover, the results demonstrate that when multiple Web applications are running in a Xen consolidated environment, not only the high CPU utilization but also the high disk utilization causes the performance degradation. The dramatic increase of L3 cache misses of the system is another cause of this performance degradation. The functions of Xen and Dom0 kernels that handle I/O operations of domains have the most contribution to this considerable increase of L3 cache misses.
