Research On I/O Performance Optimization In CPU Shared Virtual Machines

With the rapid development of cloud computing, more and more people began to study the virtualization technology in cloud computing. In the traditional environment, CPU is exclusive physical device of the operating system, while in the virtual environment, as a common phenomenon, it is shared by multiple virtual machines in order to reduce the cost of equipments. With the increase of the number of virtual machines sharing one physical CPU, there are more and more tasks in the CPU run queue, which lead to the increase of the average access latency for each task. As the virtualization technology that has been widely used, Xen’s default scheduler is credit scheduler. However, the type of masks in virtual machines is unknown to the scheduler, and I/O-intensive and CPU-intensive masks are treated the same. For I/O-intensive tasks, this latency can have a significant negative impact on I/O performance.According to the problem discussed above, this dissertation proposed an accelerate-core model which accelerates I/O processing by offloading I/O processing to a designated core. By modifying the virtual machine scheduler, each virtual machine is assigned a virtual accelerate-CPU. By designing the new credit allocation algorithm, each virtual machine can get enough CPU resources in virtual machine operating system. By modifying the process scheduling rules and increasing the I/O buffer size,the context switching frequency and the packet loss rate is reduced. By designing the ACK generation algorithm of TCP package, the performance of TCP is improved. Based on the above method, an accelerate-core model is implemented in Xen. Tests show that the model can effectively improve the I/O performance of virtual machines.