The issues of content placement and content replication for video-on-demand streaming over cellular networks are addressed in this study. Using many replications of a relatively small number of the most popular items a significant performance improvement can be achieved. Our method was verified using real video streaming data taken from traces of live content distribution networks. Simulation results show that replicating a relatively small number of video files can significantly reduce the incoming bandwidth from the Internet backbone, as well as the (time) latency for content delivery. The proposed scheme is particularly suitable for IP-based TV services, for which the content popularity can be very often predicted with relatively high accuracy. In addition, we propose a hybrid cache management scheme, in which the cache is partitioned into two components. The first component is for long-term items, and it is updated relatively rarely, while the second component is updated more frequently.