|Title||Modeling and analysis of software rejuvenation in a server virtualized system with live VM migration|
|Publication Type||Journal Article|
|Year of Publication||2013|
|Authors||F Machida, DS Kim, and KS Trivedi|
|Pagination||212 - 230|
As server virtualization is used in a number of IT systems, the unavailability of virtual machines (VM) on server virtualized systems becomes a significant concern. Software rejuvenation is a promising technique for improving the availability of server virtualized systems as it can postpone or prevent failures caused by software aging in both the VM and the underlying virtual machine monitor (VMM). In this paper, we study the effectiveness of a combination of VMM rejuvenation and live VM migration. When a VMM needs to be rejuvenated, the hosted VMs running on the VMM can be moved to another host using live VM migration and continue the execution even during the VMM rejuvenation. We call this technique Migrate-VM rejuvenation and construct an availability model in the stochastic reward net for evaluating it in comparison with the conventional approaches; Cold-VM rejuvenation and Warm-VM rejuvenation. The designed model enables us to find the optimum combinations of rejuvenation trigger intervals that maximize the availability of VM. In terms of the maximum VM availability, Migrate-VM rejuvenation is potentially the best approach. However, the advantage of Migrate-VM rejuvenation depends on the type of live VM migration (stop-and-copy or pre-copy) and the policy for migration back to the original host after VMM rejuvenation (return-back or stay-on). Through numerical examples, we show that ''pre-copy'' live VM migration is encouraged rather than pure ''stop-and-copy'' migration and it is better to return back VM to the original host soon after the VMM rejuvenation (i.e., ''return-back'' rather than ''stay-on'' policy) for highavailability. The effect of the VMM rejuvenation technique on the expected number of transactions lost is also studied by combining the availability model with an M/M/1/n queueing model. © 2012 Elsevier B.V. All rights reserved.
|Short Title||Performance Evaluation|