Signal Processing and Speech Communication Laboratory
hometheses & projects › Predicting the Latency of MQTT Brokers Using Deep Learning

Predicting the Latency of MQTT Brokers Using Deep Learning

Master Thesis
Announcement date
01 Oct 2018
Andreas Wurm
Research Areas


MQTT is one of the most widely used protocols in the Internet of Things. The performance of this protocol can be tested using statistical model checking integrated into a property- based testing tool. This approach utilizes a cost model to get predictions of the latency of the system-under-test. Multiple linear regression is used for the creation of the cost model at the moment. In this work we replace the current cost model with deep learning methods. We analyze different datasets from various experiments and broker implementations in order to find the most suitable (recurrent) neural network. We compare different architectures of standard neural networks and gated recurrent units and evaluate the models on datasets created by the test system. We show that the results of the predictions improve significantly compared to the multiple linear regression. In particular, we obtain an R 2 value of 0.9152 compared to a value of 0.8697 with the current cost model, using a dataset created with limited CPU resources. Additionally, the effort of preprocessing, data cleansing and human interactions can be lowered to a minimum. Compared to the mean latencies of the system-under-test, a simulation on the model can achieve a speedup by a factor of up to 500.