Agile auto scaling for supporting large scale cloud service platform

As most of the auto scaling algorithms have drawbacks such as latency, coarse-grained and high overhead, inspired by cache ideology, suspended virtual machines were introduced to speed the provision speed. An agile auto scaling algorithm was designed based on auto-regressive and moving average(ARMA) with two level prediction to achieve fine-grained resource allocation, and strategies such as quantile statistic, allocation of extra resources, delayed release of resources were adopted to further satisfy the quality of service (QoS). Experimental results show that auto scaling can further save cloud resources, while improving the service quality with the workload of NETEASE Cloud Reader.