Applications have different requirements in terms of bandwidth and delay to deliver the expected quality of service. For this and other reasons, new paradigms have emerged, such as software-defined networks, which allow the development of new applications for dynamic programming of routing devices in the network through a device called controller. In this work, five machine learning models for classifying network traffic are compared, which when implemented in the controller make it possible to discover the types of applications and their quality of service requirements. The obtained results demonstrate that the Random Forest model has an excellent performance in the classification of network traffic using few packets and achieves very high accuracy margins. On the other hand, it was noted that machine learning models based on streaming data are very sensitive to the dataset and the amount of packets used in the classification.