A Machine Learning-Based Framework for Building Application Failure Prediction Models

Abstract

In this paper, we present the Framework for building Failure Prediction Models (F <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> PM), a Machine Learning-based Framework to build models for predicting the Remaining Time to Failure (RTTF) of applications in the presence of software anomalies. F <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> PM uses measurements of a number of system features in order to create a knowledge base, which is then used to build prediction models. F <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> PM is application-independent, i.e. It solely exploits measurements of system-level features. Thus, it can be used in differentiated contexts, without the need for any manual modification or intervention to the running applications. To generate optimized models, F <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> PM can perform a feature selection to identify, among all the measured system features, which have a major impact in the prediction of the RTTF. This allows to produce different models, which use different set of input features. Generated models can be compared by the user by using a set of metrics produced by F <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> PM, which are related to the model prediction accuracy, as well as to the model building time. We also present experimental results of a successful application of F <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> PM, using the standard TPC-W e-commerce benchmark.