Abstract: The public switched telephone network (PSTN) provides ubiquitous availability and very high scalability of more than a million busy hour call attempts per switch. If large carriers are to adopt Internet telephony, then Internet telephony servers should offer at least similar quantifiable guarantees for scalability and reliability using metrics such as call setup latency, server call handling capacity, busy hour call arrivals, mean-time between failures and mean-time to recover. This thesis presents a reliable, scalable and interoperable Internet telephony architecture for user registration, call routing, conferencing and unified messaging using commodity hardware. The results extend beyond Internet telephony to encompass multimedia communication in general. The architecture presented in this thesis deals with two aspects: at least PSTN-grade reliability and scalability of the Internet telephony servers, and interoperable Internet telephony services such as conferencing and voice mail using existing protocols. We describe the architecture and implementation of our Session Initiation Protocol (SIP)-based enterprise Internet telephony architecture known as Columbia InterNet Extensible Multimedia Architecture (CINEMA). It consists of a SIP registration and proxy server, a multi-party conferencing server, a gateway for interworking SIP with ITU's H.323, an interactive voice response system and a multimedia mail server. CINEMA provides a distributed interoperable architecture for collaboration using synchronous communications like multimedia conferencing, instant messaging, shared web-browsing, and asynchronous communications like discussion forum, shared files, voice and video mails. It allows seamless integration with various communication means like telephone, IP phone, web and electronic mail. We present two techniques for providing scalability and reliability in SIP: server redundancy and a novel peer-to-peer architecture. For the former, we use DNS-based load sharing among multiple distributed servers that use backend SQL databases to maintain user records. Our two-stage architecture scales linearly with the number of servers. For the latter, we propose a peer-to-peer Internet telephony architecture that supports basic user registration and call setup as well as advanced services such as offline message delivery, voice mail and multi-party conferencing using SIP. It interworks with server-based SIP infrastructures.