Abstract: Through tracing the log data of four typical streaming servers of CERNET, the characteristics of streaming media and its access pattern to attempt to give evidence of the key component design or simulation in streaming system such as streaming proxy in CDN were analyzed. The results show that the size and playback duration of media files fit the lognormal distribution well but the average bit rate varies in a "bi-peak" model; references to streaming media present stronger skew popularity than Zipf distribution; temporal locality of the reference is unclosed. Both the locality and temporal correlation that exists in the request stream are enclosed that it is the hot references to the few front media files of the file rank that cause the locality, while these requests present stronger temporal correlation. Implications of above conclusions in analyzing, designing and implementing the streaming cache proxy are elucidated.