Neural Network-Based Approaches for Predicting Query Response Times

Abstract

Query response time prediction is an important and challenging problem in database systems. Especially for applications which handle large amounts of data or where time loss and deadlocks are hardly tolerated, it is very useful to predict the query response times before actual execution. This paper aims to predict query response times automatically using neural network-based approaches, and compares these approaches in terms of training time and accuracy. We implemented three methods based on artificial neural networks, and compared these methods using the TPC-DS benchmark database on Microsoft SQL Server. This study shows that two of our methods, multilayer perceptron with back-propagation and small-world network methods, present accurate results in predicting query response times within acceptable training times.