Forecasting models for self-adaptive cloud applications: A comparative study

With the introduction of autoscaling, clouds have strengthened their position as self-adaptive systems. Nevertheless, the reactive nature of the existing autoscaling solutions provided by major Infrastructure-as-a-Service (IaaS) cloud services providers (CSP) heavily limits the ability of cloud applications for self-adaptation. The major reason of such limitations is the necessity for the manual configuration of the autoscaling rules. With the evolution of monitoring systems, it became possible to employ the data-driven approaches to derive the parameters of scaling rules in order to enable the autoscaling in advance, i.e. the predictive autoscaling. The change in the amount of requests to microservices could be considered as a reason to adapt the virtual infrastructure underlying the cloud application. By forecasting the amount of requests to cloud application, it is possible to estimate the upcoming demand to replicate the microservices in advance. Hence, anticipation of the demand on the cloud application helps to evolve its self-adaptive properties. In the scope of the paper, the authors have tested various extrapolation models on the real anonymized requests time series data for 261 microservices provided by the industry partner Instana. The tested models are: Various seasonal ARIMA models with GARCH modifications and outliers detection, exponential smoothing models, singular spectrum analysis (SSA), support vector regression (SVR), and simple linear regression. In order to evaluate the accuracy of these models, an interval score was used. The time required to fit and use each model was also evaluated. Comparative results of this research and the classification of forecasting models based on the interval accuracy score and model fitting time are provided in the paper. The study provides an approach to evaluate the quality of forecasting models to be used for self-adapting cloud applications and virtual infrastructure.