In this paper, we present a new multicast architecture and the associated multicast routing protocol for providing efficient and flexible multicast services over the Internet. Traditional multicast architectures construct and update the multicast tree in a distributed manner, which causes two problems: first, since each node has only local or partial information on the network topology and group membership, it is difficult to build an efficient multicast tree; second, due to lack of the complete information, broadcast is often used when transmitting control packets or data packets, which consumes a great deal of network bandwidth. In the newly proposed multicast architecture, a few powerful routers, called m-routers, collect multicast-related information and process multicast requests based on the information collected. m-routers handle most of multicast related tasks, while other routers only need to perform minimum functions for routing. m-routers are designed to be able to handle simultaneous many-to-many communications efficiently. The new multicast routing protocol, called Service Centric Multicast Protocol (SCMP), builds a dynamic shared multicast tree rooted at the m-router for each group. The multicast tree can satisfy the QoS constraint on maximum end-to-end delay and minimize tree cost as well. The tree construction is performed by a special type of self-routing packets to minimize protocol overhead. Our simulation results on NS-2 demonstrate that the new SCMP protocol outperforms other existing protocols and is a promising alternative for providing efficient and flexible multicast services over the Internet.