Аннотация:To achieve full virtualization support on Xen, Intel extended the Xen project to support unmodified guest OSs using Intel virtualization technology. In this paper, we describe the performance tuning work in the process of enabling full virtualization on Xen. Firstly, we briefly summarize our extensions on performance tuning tools. Then, through data analysis collected by the performance tuning toolkits, we observe that system performance bottlenecks are mainly introduced by memory and I/O virtualization overheads. Finally, we describe our optimizations and possible proposals to address these bottlenecks for resolving performance bugs in Xen/VT architecture. In that process, more specifically, we improved the virtual platform performance by 10% for compute-intensive benchmarks, by moving virtual PIC to hypervisor. In the meantime, we enhanced virtual IDE disk DMA operation performance through concurrent processing and increased virtual network interface card performance more than ten times by reconstruction using an event-driven mechanism. In addition, to achieve higher quality in system X, we also optimized a virtual video graphic card, and several times we got performance gain through our shared virtual video memory approach.
