Cloud autoscaling simulation based on queueing network model

For the development of a predictive autoscaler for private clouds, an evaluation method was needed. A survey of available tools was made, but none were found suitable. The CloudAnalyst distribution of CloudSim was examined, but it had accuracy and speed issues. Therefore, a new method of simulation of a cloud autoscaler was devised, with a queueing network model at the core. This method's outputs match those of a load test experiment. It is then evaluated with basic threshold-based algorithms on traces from e-commerce websites taken during Christmas. Algorithms based on utilization, latency, and queue length are assessed and compared, and two more algorithms combining these metrics are proposed. The combination of scaling up based on latency and down based on utilization is found to be very stable and cost-efficient. The next step will be the implementation of predictive methods into the autoscaler, which were already evaluated in the same R language environment. (C) 2016 Elsevier B.V. All rights reserved.