Research On Software Rejuvenation For Virtualized Systems

The virtualized system resently plays a vital role in cloud computing for less dependency on hardwares along with easier maintenance,better load balance and power management.Similar as other software,VMs and underlying VMM in virtualized systems also experience aging issues.In the last years,software rejuvenation attracts significant attention for its proactive fault management principle which targets to reduce software aging risks.This dissertation mainly focuses on researches on software rejuvenation of virtualized systems,including rejuvenation analytical models and VM live migration of migrate-VM rejuvenation.The major tasks and research achievements of this dissertation include:1)To counteract software aging for a single-server virtualized system,it introduces four rejuvenation analytical models with different rejuvenation policies for the VMs and underlying VMM,then analyzes the models by Markov regeneration theory and proposes the optimum rejuvenation scheduling algorithm.The experiment results show that the optimal rejuvenation scheduling can bothimprove system steady-state availability and reduce system downtime cost.2)Regarding to the strong dependency of VMs and underlying VMM among multiple servers,after the detail comparison of time based rejuvenation models for cold-VM rejuvenation,warm-VM rejuvenation and migrate-VM rejuvenation,a time and load based rejuvenation analytical model in Stochastic Reward Net(SRN)is proposed.The experience result shows that migrate-VM rejuvenation has higher system steady-state availability and less number of transaction loss than traditional cold-VM rejuvenation and warm-VM rejuvenation,and the time and load based rejuvenation analytical model is better than the time based model in respect of system availability and throughput rate,and is more stable in face of the dynamic change of system load.3)To reduce the rejuvenation cost of VM migration in multi-servers' virtualization system,an optimization of memory state transfer(MSTO)in the iterative copy stage of the pre-copy approach is proposed.MSTO designs an adaptive transfer window and dynamic migration candidate pages selection policy migration in accordance with temporal locality and spatial locality principles,so as to ensure dirty pages and link bandwidth are well suited during whole migration duration.Experiment results show the optimized approach is better than the built-in pre-copy based approach of Xen in terms of downtime,total migration time and the number of transferred pages.