Research And Performance Optimization Of A Hardware Virtualization Machine Xen

With the fast growing of Internet and the emerging requirements for service consolidating, virtual host and resource management, the virtual machine (VM) technology is experiencing resurgence. In the mean time, VM on x86 has an even more bright perspective due to more and more wildly deployment of x86 server. x86 platform itself, however, doesn't provide a strong support for VM management. Currently, there are two popular x86 VM implementations: full-virtualized VMware and para-virtualized Xen. The former can run most mainstream OS, while it brings quite performance penalty. The latter needs modify a few hardware related code of OS. With OS cooperating with VM monitor, it runs excellent in light of performance evaluation. But it has its own shortcoming: it can't deploy Windows which is widely used on Desktop but closed-sourced.In response to market request, Intel released VT-x, a technology to support VM inside the chip, which is aiming to provide more efficient VM management and less cost. Based on Xen's advantages of open sourcing and excellent performance, Intel Linux VMM team add to it VT-x support and bring a VMX VM model. The project's objective is a high performance full-virtualized implementation, which can overcome respective disadvantages of VMware and Xen.The author participated in the project during her internship in intel, mainly focusing on performance test and optimization. We found as VM introduces more management overhead, it may result in dramatic performance difference between virtual machine and native environment. Currently there are few efficient ways to tune their performance, except utilizing those general benchmark tools like SPEC CPU2000, SPEC JBB2000, lmbench, TTCP, etc. Then we studied our VM model's characteristic, developed Xentrace a, which is a low level Xen virtual machine performance tuning tools. We used it to collect and parse data of virtual machines with different workloads. Looking into the Xentrace's data, we firstly addressed a design defect, and then successfully resolved a series of critical virtualized I/O performance issues, which is the main cost of VM. Benchmark scores showed the changes really boosted VM's performance.