Machine Learning for Performance Prediction of Data Distribution Service (DDS)

Networking middleware following the Data Distribution Service (DDS) specification is used in real-time mission-critical systems such as autonomous vehicles, energy management systems, and air traffic control. DDS follows the publish-subscribe communication pattern and offers a set of Quality of Service (QoS) parameters, allowing the users to align the data communication to the needs of the application.Configuring DDS to achieve the required performance is a difficult task, given the large space of QoS parameter values. Experimental evaluation of performance levels with a real DDS system for different QoS configurations can be complex and require substantial time and resources.We propose the use of Machine Learning (ML) models to predict the performance metric distribution of DDS under different configurations. This is done by using performance measurements of some configurations to train an ML model. The trained model can then be used to predict the performance distribution 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. To the best of our knowledge, this is the first time this approach has been applied to DDS performance evaluation.We used random forests (RF) as an ML method and linear regression (LR) as a baseline. We selected thirteen performance metrics, and for each, we trained an RF model and tuned its hyperparameters. We tested the final models on system configurations unseen during training, both for parameter values within the training range (interpolation) and outside the training range (extrapolation).The RF models show better predictive accuracy than the LR baseline. This paper focuses on the models for throughput and latency - the two well-established performance metrics. The models demonstrate coefficients of determination greater than 0.9 and 0.8, respectively, for different unseen system configurations in interpolation, but work less well in extrapolation cases.We conclude that the proposed ML models offer a way of predicting the performance distribution of a range of configurations when interpolation is used. Since model prediction is computationally much cheaper than relying on experimentation, it is a useful tool to guide DDS system parametrisation and design.