On-demand video delivery with a Content Delivery Network (CDN) solely based on set-top-boxes has been recently introduced. In previous works, videos are separated into pieces and randomly stored in boxes, requests from clients being redirected to the nearest boxes. However, random strategy may lead to some costly and inefficient allocation. For example, a video part could be placed much further than other parts of the same video, while two close boxes host the same part. This paper aims to reduce downloading cost by exploiting the network location of boxes when allocating the videos. We show that optimizing the allocation is NP-hard. In this paper, in order to analyze the impact of different content allocation methods on cross-domain traffic, we compare an existing approximate algorithm and two simple heuristics with random allocation. Simulations with realistic network settings demonstrate that inter-AS traffic can be efficiently reduced, when knowledge on network topology is integrated in the allocation algorithm, such as links and relationships among Autonomous Systems.