Research On Rolling Based Cloud Service On-line Upgrade Technique

Applications are undergoing a big change in the way of how they are developed and deployed, shifting from desktop and enterprise software toward web and cloud computing services. Traditional operation way of “testing-release-maintenance” is ill-suited for today’s Internet environment, due to the extremely rapid changing pace of resources and user demands. On the contrary, current “monitoring-diagnosisadjustment” practice becomes more and more popular. As the most representative methods of fulfilling “adjustment” process, software upgrading activities are experiencing a frequent interval, sometimes the period between two releases even reach days of time. In order to maintain high availability under this circumstance, the important significance of online upgrading techniques has been highlighted, and it attracts researchers’ attention.In practice, most cloud service adopts “multi-instances” way of deploying, aiming at a better elastic ability to handle unknown workloads. This clustering structure opens a new opportunity for cloud service online upgrading problem solution. Not like traditional dynamic software upgrades, which focus on upgrading single-node systems on the fly, the characteristics of clustering allow components or instances to be temporarily inaccessible while the whole service is still been provided normally. In this paper, we mainly made three contributions related to cloud service online upgrading problem relies on this assumption.Firstly, we proposed rolling based online upgrading schema, and developed a prototype system implementing this schema based on GlassFish platform. By disabling, upgrading and rebooting some instances at a time according to certain policies, we achieved our goal of upgrading the whole service with minimal loss of availability.Secondly, in order to solve mixed version problem, and the service capacity loss problem along with the solution of mixed version, we presented delayed switch mechanism. By dividing the upgrade process into two phases, it logically isolates the different versions as well as reserves as many instances as possible. Experiments with a typical e-commerce service named Rubis have been conducted. The results show that we achieve above 99% service availability during upgrading process, and improved the average responsible speed to 1.4 times of the origin one.Thirdly, we have designed Task Status Perception mechanism, an effective method of finding the exact upgrading time for each service instance. By adding a task status perception queue for each instance at the load balancer side, we can get the exact time for upgrading the instance by judging the corresponding queue length. The experiments have been tested under current on-line upgrading methods include rolling upgrade, split mode upgrade and delayed switch upgrade. All results confirmed the effectiveness and the efficiency of this algorithm.