Joint Tracking of Performance Model Parameters and System Behavior Using a Multiple-Model Kalman Filter

Zhen ZHANG  Shanping LI  Junzan ZHOU  

IEICE TRANSACTIONS on Information and Systems   Vol.E96-D   No.6   pp.1309-1322
Publication Date: 2013/06/01
Online ISSN: 1745-1361
DOI: 10.1587/transinf.E96.D.1309
Print ISSN: 0916-8532
Type of Manuscript: PAPER
Category: Software Engineering
performance modeling,  tracking filter,  multiple-model method,  queueing theory,  resource management,  

Full Text: PDF(762.8KB)>>
Buy this Article

Online resource management of a software system can take advantage of a performance model to predict the effect of proposed changes. However, the prediction accuracy may degrade if the performance model does not adapt to the changes in the system. This work considers the problem of using Kalman filters to track changes in both performance model parameters and system behavior. We propose a method based on the multiple-model Kalman filter. The method runs a set of Kalman filters, each of which models different system behavior, and adaptively fuses the output of those filters for overall estimates. We conducted case studies to demonstrate how to use the method to track changes in various system behaviors: performance modeling, process modeling, and measurement noise. The experiments show that the method can detect changes in system behavior promptly and significantly improve the tracking and prediction accuracy over the single-model Kalman filter. The influence of model design parameters and mode-model mismatch is evaluated. The results support the usefulness of the multiple-model Kalman filter for tracking performance model parameters in systems with time-varying behavior.