Performance Analysis And Enhancement Of Virtualization On Multi-core Processors

Multi-core is a processor which integrates two or more complete computing en-gine. It has become an important technology fulcrum of server technology. The futuretrend of multi-core processors is heterogeneity. Dynamical heterogeneity makes agreater challenge for the future operating system. Intermittent fault is a common phe-nomenon in hardware, which means hardware errors burst during a period of time.Intermittent faults are caused by ?uctuation of voltage, temperature or some other fac-tors. Since the factors causing intermittent fault are various, the duration of burstyfault have a wide timescale.Virtualization is a concept closely related to multicore processors. The basicthought is that a unique physical resource may act like multiple virtual resources whichcan be operated separately. Obviously, multicore processors let the implementation ofvirtualization to be more easy, because each core can process different things. In re-cent years, Xen has already become one of the most popular virtualization technology.Xen is an open source virtual machine monitor for IA-32 (x86, x86-64), IA-64 andPowerPC 970 architectures. Xen allows several operating systems to share ordinaryhardware in a safe and managed manner, without sacrificing performance and func-tionality. One of the core features of Xen is multiple tasks. The hypervisor mustpromise each running virtual machine to get some CPU time. Similar to scheduling inoperating system, scheduling in Xen is a tradeoff between running each domain fairlyand achieving good overall throughput.The scheduling in operating systems have a significant impact on performanceand power in dynamically heterogeneous multi-core processors. Similarly, CPU schedul-ing algorithm that lacks of considering dynamic heterogeneity of multi-core is likelyto reduce the virtual machine performance and reliability. In a dynamic heterogeneous multi-core processors, the operating system scheduling efficiency will have a signifi-cant impact. Since intermittent fault may in?uence multi-core processors, it certainlymay impact upon scheduling in Xen. We create a timing model which simulates CPUintermittent fault in Xen hypervisor. In the model, domains in the unmodified Xensystem crash immediately.We analyzed the algorithm of Credit Scheduler, the default Scheduler in recentversions of Xen, and found that the scheduler lacks the consideration of heterogeneousfactors, especially intermittent fault. Once bursty faults occur on a certain physicalCPU, the Xen system may quickly crash. The basic reason is that the the scheduler isunconcerned about the change of CPU status. If a CPU is suddenly of?ine, runningand waiting threads(In Xen, they are referred as virtual CPU) on this CPU will beterminated. Therefore the throughput of the system may fall off promptly and terribly,or even the whole system may crash.In this paper, we propose an adaptive strategy for scheduling in Xen, in order toeliminate the in?uence of intermittent faults. We trace the status of CPU periodicallyin the scheduler. Once a CPU is found from online status to of?ine status, we manageto migrate all loads on the faulty CPU to other CPUs. The experiment shows thatwhile intermittent fault happens frequently, the Xen system with our strategy can stillmove on, and the performance reduces smoothly. Meanwhile, the overhead of tracingis quite low.