A Reliable Multicast Framework for Light-weight Sessions and Application Level Framing

S. Floyd, V. Jacobson, S. McCanne, C-G Liu, L. Zhang, "A Reliable Multicast Framework for Light-weight Sessions and Application Level Framing," ACM SIGCOMM Conference, (August 1995).

This paper talked about SRM (Scalable Reliable Multicast), a reliable multicast protocol for application level framing and light-weight sessions. This is based on the IP Multicast group delivery model. The paper starts off by identifying some key problems in extending basic TCP style multicast such as ACK implosion, responsibility of maintaining dynamic group membership information, and the whole concept of single receiver based RTT. Keeping this problems in mind, the authors proposed that any sender based reliability scheme like TCP wouldn't really work in such a scenario and what was needed was a receiver based scheme. Further, the reliability should be maintained on a 'application data unit' (ADU) basis. This helps in putting the application in control and also since ADU names can be independent of sending hosts, it can use the anonymity of IPMulticast to exploit redundancy of multiple receivers.

The entire system was based on the wb framework which was a network conferencing tool. The underlying idea was to prevent duplicate requests for the same ADU by various receivers in the multicast group. This is done by promiscuously hearing retransmit request and responses by other multicast receivers. Whenever a receiver detects a loss, it schedules a repair request for a random time in future between [Ci*d, (C1+C2)*d] where d is the the receiver's estimate of the one-way delay to the host. Similarly, when host B receives a request from A that B is capable of answering, B sets a repair timer to a value between [D1*d, (D1+D2)*d]. The values of C1, C2, D1, D2 are set such that there is as less retransmit requests as possible in case of lost packets.

The authors showed examples of various topologies such as chain, stars etc. The simulations were done on random trees as well as large bounded-degree trees to show the scalability of the protocol. They also proposed an adaptive adjustment of the constants by means of random timer algorithm that depended on average duplicate requests. Further, to conclude the paper, the authors talked about how is this protocol generic in nature and is not just limited to wb style multicast model.