Pitakrat, Teerat and van Hoorn, Andre (2014) Investigating the Use of Bayesian Networks in the Hora Approach for Component-based Online Failure Prediction. [Talk] In: Symposium on Software Performance 2014: Joint Descartes/Kieker/Palladio Days. , November 26-28, 2014, Stuttgart, Germany .

Preview

Slideshow 19_Pitakrat_Hora-slides.pdf - Presentation Download (1MB) | Preview

Abstract

Online failure prediction is an approach that aims to predict potential failures that can occur in the near future. There are a number of techniques that have been used, e.g., time series forecasting, machine learning, and anomaly detection. These techniques are applied to the data that can be collected from the system and that contain information regarding the current state of the system, such as, response time, log files, and resource utilization.

The existing works which employ these prediction techniques to predict failures can be grouped into two categories. The first category approaches the task by using the technique to predict failures at specific locations in the system. For example, time series forecasting may be used to predict the response time at the system boundary. Once it is predicted that the response time tend to go beyond a certain value in the near future, a warning is then issued. On the other hand, the second category applies the prediction technique to the whole system, i.e., using the data collected from all locations and creating a model that can analyze and conclude from these data whether a failure is expected on the system level.

In our work, we take another direction by combining techniques in the first category with the architectural model of the system to predict not only the failures but also their consequences. In other words, the existing prediction techniques provide prediction results of each component while the architectural model allows the predicted failures to be extrapolated and further predicts whether they will affect other components in the system.

We have developed the prediction framework based on Kieker’s pipe-and-filter architecture and employed its monitoring capability to obtain the data at runtime. The architectural model of the system is extracted from the monitoring data and used to create a Bayesian network that can represent the failure dependency between components. The prediction results obtained from each component failure predictors are forwarded to the Bayesian network to predict the failure propagation.

Actions (login required)

View Item