The future Internet scenario consists of a higher number of users and applications, which demand more resources from the communication infrastructure. Techniques for providing performance and scalability, such as Traffic Engineering (TE), will always be necessary even if the transmission rate is very high, because of such demands. Quality of Service is one of the solutions that can be used to improve the traffic engineering in the Internet, with the most referenced architecture: DiffServ. In general, TE needs traffic classification to accurately identify the input traffic and manage it properly. However, the current DiffServ port traffic classifier is considered outdated. This paper presents a performance evaluation of machine learning traffic classification solutions applied to DiffServ, and investigates their benefits on network performance. For a backbone network with 40 nodes, the performance of the network can increase up to 15% for both data and voice traffic.