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

The reliable multicast framework paper is interesting in it's development of their model. They begin with an application (wb, a multiuser whiteboard) and from there attempt to factor out what they deem traits that could be common to many applications that need multicast. Their overall approach is to keep as much functionality as they can in the client (as dictated by the end-to-end argument), doing only best effort delivery and requiring the clients to implement reliability and packet ordering enforcement.

The key contribution of this paper seems to be their mechanism for retransmitting lost packets. These losses are indicated by the clients, and the idea is to minimize the amount of extra data that needs to be sent. Clients determine a lost packet by their reception of a packet with a higher sequence number than expected. If this happens, a timer is set based on the distance of the client from the source (further away = longer timer). When this timer expires a request for retransmission is sent. If the packet is received while waiting, the timer is canceled. In this way it is expected that the first host after a lossy link will request retransmission and that it will get a prompt reply from the host just upstream of the loss. A small amount of randomness is added to the timer to try and still only use one request when two hosts equidistant from the source both miss a packet. In theory the randomness will allow one host to request retransmission first and since the response is multicast the other host will see it as well and cancel its timer.

This means hosts need to cache old requests so they can respond to retransmission requests. The amount of state needed to make this all work is not completely clear, nor is the issue addressed in the paper. Still, it's an interesting approach, and the authors run a number of simulations showing that if all state is kept their protocol performs well.