Ad-hoc overlay networks are increasingly used for sharing static bulk content but their promise for scaling the delivery of on-demand, realtime content is yet to be tapped. In this paper, we show that overlay networks could be used ef.ciently to distribute popular real-time streaming media on-demand to a large number of clients. We propose and evaluate OSMOSIS a cache-and-relay end-system multicast approach, whereby a client joining a multicast session caches the stream, and if needed, relays that stream to neighboring clients which may join the multicast session at some later time. OSCMOSIS is fully distributed, scalable, and ef.cient in terms of network link costs. We present analytical and empirical results of our evaluation of OSMOSIS. Our analysis establishes OSMOSIS scalability characteristics under a variety of assumptions. Our simulations are over large, synthetic random networks, power-law degree networks, and small-world networks (all of which could well be representative of ad-hoc overlay topologies, as well as over large real router-level Internet maps.