BitTorrent (BT) is an effective tool for large-scale P2P content services in the face of dynamic nature of Internet. A P2P content service is usually characterized by the fact that its users leave and join the service frequently and are self-interested. It is further complicated by the flash crowd phenomenon. However, BT does not perform well in streaming applications in which contents are supposed to be delivered timely. For example, users of streaming video expect to start viewing the video as early as possible. This is unlike a BT user who usually has to wait until the entire content is downloaded before one could start to view the video. Furthermore, each byte of the streaming video must arrive in time to avoid unpleasant playback jitters. In this paper, we present our design of USVoD piece selection mechanisms into BT to efficiently overcome the above problems. Instead of selecting rarest pieces, our mechanism selects pieces based on the following principles, (1) moving-window strategy, i.e., selects all pieces which will be played back within a fixed time interval in the near future, (2) on-line cache mechanism, i.e., randomly selects a certain pieces which will be used after the moving window interval, (3) uniform pre-sampling cache mechanism, i.e., in addition to pre-fetching initial portion of the video, it also pre-fetches some randomly selected pieces into the specific cache before playback, and (4) some useful modifications to BT. We evaluate the performance of USVoD and BiToS on a mixture of Poisson arrival and flash crowd by using tradeoff between initial delay and delay jitter as evaluation criterion. Simulation results show that USVoD significantly outperform BiToS.