Combining time series prediction models using genetic algorithm to autoscaling Web applications hosted in the cloud infrastructure

摘要

In a cloud computing environment, companies have the ability to allocate resources according to demand. However, there is a delay that may take minutes between the request for a new resource and it being ready for using. This causes the reactive techniques, which request a new resource only when the system reaches a certain load threshold, to be not suitable for the resource allocation process. To address this problem, it is necessary to predict requests that arrive at the system in the next period of time to allocate the necessary resources, before the system becomes overloaded. There are several time series forecasting models to calculate the workload predictions based on history of monitoring data. However, it is difficult to know which is the best time series forecasting model to be used in each case. The work becomes even more complicated when the user does not have much historical data to be analyzed. Most related work considers only single methods to evaluate the results of the forecast. Other works propose an approach that selects suitable forecasting methods for a given context. But in this case, it is necessary to have a significant amount of data to train the classifier. Moreover, the best solution may not be a specific model, but rather a combination of models. In this paper we propose an adaptive prediction method using genetic algorithms to combine time series forecasting models. Our method does not require a previous phase of training, because it constantly adapts the extent to which the data are coming. To evaluate our proposal, we use three logs extracted from real Web servers. The results show that our proposal often brings the best result and is generic enough to adapt to various types of time series.