Automatic scaling of machine resources in complex computing systems

Abstract

This thesis examines the feasibility of detecting future queues in complex computing pipelines using historic time series data as training data for a recurrent neural network. It is suspected that surges of information that will be processed at different stages in the system spread and affect other processes. By predicting how large the queues are going to be a few minutes in the future, preemptive measures can be taken in order to mitigate the spikes in workload. This can be done by scaling the computing power at every node accordingly ahead of time. In order to find the useful information patterns in a very large feature space, different feature selection methods are tried and evaluated. It is found that choosing features based on their relevance to the target feature performs better than choosing features that span the feature space. Three different ways of looking at the results are tested: Naive prediction of future queue sizes, WTTE-RNN and the Sliding Box model. It is found that some predictive power exists in the former two, while the Sliding Box model performs poorly, but more tuning and data collection is needed before putting the results into production.