MA Abschlussvortrag, Benjamin Schindele, Ort: O27/545, Datum: 09.03.2017, Zeit: 11:15 Uhr

Current database monitoring systems often lack the ability to forecast the future perfor¬mance of database systems. As a consequence, database administrators can often only react to performance problems, instead of avoiding them in the first place. We propose a new strategy for predicting the prospective performance of a database system. We suggest, that the systems performance depends an future values of certain database and server metrics, like CPU usage or Batch Requests executed per second. These values can be predicted by applying time series forecasting techniques. Usually sta¬tistical models are used to perform predictions of time series, however in recent years machine learning models have become serious alternatives. We therefore employ three different machine learning models, namely feedforward neural networks, support vector machines and gaussian processes, and conduct an evaluation of their forecast accuracy with historical performance data of a database system. The evaluation results are then used to design a forecasting system for database performance. Additionally, a prototype with basic forecasting capabilities is implemented as a proof of concept.