Online traffic classification is getting into troubles when the network applications are exploding. The traditional machine learning methods based on statistical characteristics of packets work well in stable network environment

but not in congestion environment with serious delay and packet loss. Therefore

a novel classification method based on granular computing is proposed in this paper. Granular computing belongs to the field of artificial intelligence computing

which is usually used to process missing

incomplete or noisy data. So we first define granules for the traffic

then construct the relations between the granules

and finally establish the relation matrix. The traditional statistical characteristics are only the special case of the relation matrix when the scale is the largest. The granular relation matrix can describe the traffic more comprehensively and classify them more accurately. The experiment results show its validity and advantages when compared with other methods.