Reliable multicast is a powerful primitive, useful for data replication, event notification (publish-subscribe), fault tolerance and other purposes. Yet many of the most interesting applications give rise to huge numbers of heavily overlapping groups, some of which may be large. Existing multicast systems scale scale poorly in one or both respects. We propose the QuickSilver Scalable Multicast protocol (QSM), a novel solution that delivers performance almost independent of the number of groups and introduces newmechanisms that scale well in the number of nodes with minimal performance and delay penalties when loss occurs. Key to the solution is a level of indirection: a mapping of groups to regions of group overlap in which communication associated with different protocols can be merged. The core of QSM is a new regional multicast protocol that offers scalability and performance benefits over a wide range of region sizes.