| With the rapid development of Internet and related streaming media technology, streamingmedia is growing rapidly. Streaming media also caused a great impact on the network managementbecause of its large volume of data, high real-time requirement. In order to achieve a better controland management on network traffic and QoS guarantee on streaming media, it is necessary toclassify streaming media effectively. One of key points about achieving an effective trafficclassification is to analyze their features and select some effective ones.According to the7charactertistics: packet size distribution, average rate, the alternating of IP,ratio of downlink and uplink, number of sub-flow’s segments,the mean value of inter-packet time,6kinds of video applications: two types of HTTP streaming (high definition and standard definition),HTTP download, live TV (e.g. Sopcast), QQ video and XUNLEI are studied in this thesis. Theauthor finds that each application has significant packet size distribution characteristics. Throughthe comparison between the different data captured at different time, it is found that the packetlength distributions of network applications are stable. After computing the Hellinger distance ofapplications, the author finds that QQ has larger distance than the other5applications. Highdefinition HTTP streaming, standard definition HTTP streaming, HTTP download have similarpacket size distribution. From the perspective of the ratio of downlink and uplink, standard and highdefinition HTTP streaming are easier to be identified. XUNLEI has the most number of sub-flowsegments. According to the analysis, two characteristics: ratio of downlink and uplink, number ofsub-flow’s segment can be used to classify the selected applications. The experimental results usingSupport Vector Machine and these two characteristics obtained higher classification accuracy. |
