City Research Online

Abstract

Data Distribution Service (DDS) is a specification of networking middleware used in real-time mission-critical systems such as autonomous vehicles, energy management systems, and air traffic control. It follows the publish-subscribe communication patterns and adopts the use of Quality of Service (QoS) parameters, allowing customisation of the data dissemination process in real-time.

When setting up DDS systems, practitioners must ensure the required performance levels are achievable by setting appropriate QoS and non-QoS parameters. The evaluation of performance levels can be done by running experimental performance tests for different QoS configurations to find a suitable or even a near-optimal system configuration. However, evaluation via measurements with real DDS systems can be complex and expensive, needing potentially substantial time and resources.

This paper introduces, to our knowledge for the first time, the use of machine learning (ML) models to predict the performance of DDS under different system configurations. This is done by testing some system configurations and using the performance measurements to train a model. The trained model can then be used to predict the performance of DDS under other system configurations. Since the prediction is computationally inexpensive, we can predict the performance of many different configurations to find a suitable one for given requirements.

As an ML method, random forests have been used in this paper and as a baseline we use a linear regression model.
We selected six performance metrics and for each one we trained a random forests model and tuned its hyperparameters. We tested the final models on unseen system configurations in interpolating and extrapolating with respect to the system parameter values. The random forests models show strong predictive performance and are significantly better than linear regression. Five of the eleven random forests models have a coefficient of determination greater than 0.8 for unseen system configurations in the extrapolation setting.

With these models it is possible to explore a much wider range of parameters than could be done with experimentation alone. We therefore believe that this approach can be beneficial for DDS system design.

Publication Type:	Report
Additional Information:	Copyright, the authors, 2023.
Subjects:	Q Science > QA Mathematics > QA75 Electronic computers. Computer science
Departments:	School of Science & Technology > Computer Science

Export

Downloads