Ultra-high speed networks require real-time traffic classification in order to identify the presence of certain network applications and utilize network resources to ensure these applications run smoothly. Machine learning provides a promising alternative for traffic classification based on statistical flow features, avoiding raising privacy and security concerns. Accurate traffic classification, however, is an expensive procedure that can increase networking latency and decrease bandwidth. As an open specification, the OpenFlow protocol provides the flexibility of programmable flow processing to perform more complicated statistical analysis. So, enhanced with machine learning algorithms and OpenFlow extensions, my research focuses on the design and implementation of traffic classification system that accurately classifies traffic without affecting the latency or bandwidth of network.