Research On Inter-domain Communication And Performance Model Of Hardware Virtual Machine

With the resurgence of virtualization technology, hardware-assisted virtualization technique was invented and widely used. With the support of Intel VT or AMD SVM, a host machine can harbor multiple guest operating systems, which are called Hardware-assited Virtual Machince (i.e., HVM for short), without the revisions conducted by para-virtualization technique. However, the I/O devices of the HVMs are emulated, which makes I/O processing complicated, and places a big challenge for research on improving the I/O performance of HVM.We choose Xen which is developed by University of Cambridge as the research platform. The thesis compares the mechanism of inter-domain communication and I/O device model of para-virtualized (i.e. PV in short) guests with HVM guests. In order to improve the I/O performance of HVM, we give an approach that improves the I/O performance by introducing the I/O handling model of PV guests into HVM guests. Experiments show that this approach can improve the I/O performance of HVM by 6.7 times, and save 50% CPU resource compared to the traditional emulated I/O devices.In the hardware-assist virtualized systems, the architecture and operational mode of computer system are changed a lot. Therefore, constructing a performance model for such systems is greatly helpful for making further study on HVM guest systems. This thesis gives a performance model and a quantitative approach of HVM which are based on the analysis of I/O device model of HVM by using the queuing network model. Moreover,we take Web server in HVM as an example, identifying the bottleneck of HVMs with different resource allocation mechanisms.With the research on the performance model of HVM and the analysis of performance bottlenecks, we give a policy for performance adjustment. Under a given SLA (Service Level Agreement), this policy can allocate the appropriate amount of CPU resources to HVM guest systems. The experiment results show that this policy allocates the exact amount of resources that can guarantee the quality of service of the guest systems by a maximum error blow 10%. By this way, it can save a lot of physical resources.