Video-on-Demand (VoD) services have experienced an explosive growth in recent years, and are currently used by millions of users [3]. In traditional server-based VoD, end-user terminals download video contents from a video provider upon user’s request. Since server-based systems are nonscalable and subject to server bottlenecks, a peer-supported approach, where some of the contents is provided by peer servers, has emerged as an effective solution for that problem. In this approach, peers (user terminals) use their storage space and upload bandwidth to replicate video content and serve the content to other peers when needed. A fundamental design issue in peer-supported VoD systems is to determine the right replica placement (allocation) strategy so as to make best use of the upload capacity of peer servers. While the literature deals with P2P replica placement, it is mainly geared towards file sharing systems. For example, optimized the network bandwidth usage in replication; while [4] maximized file availability. These works pay little attention to the bandwidth limitation of peers, a key concern for VoD. To this end, we consider a multi-region peer-supported VoD service model where peer servers are located in different regions, and serving a video across regions is more expensive than within a region. The demand distribution can vary across the regions. Service can be granted by dedicated servers as a fall back, but at higher cost. The system operations requires taking care of two tasks, which are best described as if done in two stages: First, video replicas are placed by the central server at the peer servers. Then, video requests submitted by the users are assigned-to and provided-by the proper peer servers. The multi-region model is a natural fit for the network of a large content provider. For example, cable operators use video hub offices in each metropolitan area (all inter-connected) to serve the local subscribers [1, 2]. The highly variable demand distribution (some videos are highly popular while very many others are lowly popular, following heavy tail distributions, see [3]) poses challenges on the system designer regarding the number of replicas that should be placed in the system; the use of a simple assignment, like the proportional mean (proved to be optimal placement under some conditions, see e.g. [5]) may be quite inefficient as shown in the following simple example: Consider a system with two video types, A and B. The de-