Research On Hybrid Cluster Scaling Method Based On Docker Container

There is a lot of changes that are brought about by the Internet generation. Enormous amount of users, various online services give companies tremendous benefits and high traffic, heavy burden of the cluster which support the online services at the same time. Many clusters are built on virtual machines even in cloud environments and their scaling methods are mainly reactive strategies based on thresholds. User requests response time become longer even if the cluster can be dynamically scaled according to the growing workload in cloud environments. And the waste made by free servers will lead unnecessary expenditure when the workload keeps at some low levels. There are two questions which companies have to think about:how clusters can be scaled to fit the changing workload while improving scaling speed and reducing scaling cost. This thesis proposes one hybrid scaling method based on Docker container aiming at these two questions above.Firstly, we investigated the research status of scaling method in both domestic and overseas, confirmed the feasibility of Docker container and virtual machines hybrid deployed cluster, summarized three theoretical factors of scaling method. We analyzed the reason of long response time and high cost. And a pricing formula was presented based on the survey of common cloud environment pricing rules.Secondly, we proposed the hybrid scaling method which combined reactive strategy based on thresholds with proactive strategy. Auto-regression Model was used to predict the coming workload since workload is one kind of time series which has self-similarity. The decision of using Docker container to scale or virtual machine depended on the workload type that can be judged by its gradient change rate. The number of servers scaling needed could be calculated using Queuing Theory Model according to the current workload and capability of the cluster.Thirdly, pseudo codes are showed which described the scaling in and out methods. Based on the design of scaling method, an auto-scaling controller is implemented that consists of monitor, modeler, alerter and scaler who can scale the target cluster automatically.Finally, for the purpose of testing the hybrid scaling method, we designed the workload generator and two typical workloads:period workload and burst workload. Testing result showed the validity of the hybrid scaling method compared with one pure reactive scaling method.