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).

One line summary: This paper describes Scalable Reliable Multicast, a minimal framework from which applications designers can build multicast functionality suitable to their application; it also describes some analytical results pertaining to request/repair algorithms in multicast, along with several simulations.

Summary

This paper describes a reliable multicast framework called Scalable Reliable Multicast (SRM). SRM builds off of the principles of Application Level Framing (ALF), which “explicitly includes an application’s semantics in the design of that application’s protocol’, and Light-Weight Sessions (LWS), which centers on a “light-weight rendezvous mechanism based on the IP multicast distribution model” with receiver-based adaption. Thus, the SRM is designed to meet only the minimal definition of reliable multicast so as to not force on applications unnecessary overhead for functionality that they do not need. It is also designed to adhere to the core principles of TCP/IP in that it only requires the basic IP delivery model and dynamically adjusts control parameters based on observed performance, much like TCP. The authors argue that receiver-based reliability is more appropriate for SRM than sender-based reliability because the fate-sharing-based coupling of unicast does not generalize well to multicast (due to such factors as the ACK implosion effect and the need for the sender to maintain state about each of the receivers in the receiver set), and because the vocabulary of unicast conventions migrates poorly to multicast. SRM attempts to serve as a skeleton common to scalable, reliable multicast applications that supply the framework with details such as a namespace, policies and mechanisms for apportioning bandwidth, etc.

The authors go on to describe a network conferencing tool that provides a distributed whiteboard called wb, which builds off of SRM. In wb, users are members, each of whom has a globally unique identifier which is used to label pages they create and edit. Wb assumes that all data has a unique name, that the name always refers to the same data, that source-IDs are persistant, that IP multicast datagram delivery is available, and that all participants join the same multicast group. The paper describes wb’s instantiation of SRM. It then describes more generally request/repair algorithms for several simple topologies, including chains, stars, and bounded degree trees. In the context of these algorithms, it defines deterministic suppression of duplicate messages, and probabilistic suppression. They find, via simulations, that their algorithm that uses fixed timer parameters performs well in random or bounded degree trees when every node in the tree is a member of the multicast group. They use this to motivate the development of an adaptive algorithm for request/repair that adjust timer parameters as a function of delay as well as of duplicate requests or repairs in recent recovery exchanges. They demonstrate trade-offs between low delay and low number of duplicates.

Critique

I didn’t really like reading this paper. I don’t really like simulations, for one. Also, that the authors chose what they admit to be arbitrary values for some of their parameters in their algorithm annoys me. Nevertheless, however simplified the topologies in their analysis section, I would probably agree that it is good to have some mathematical results for many problems, in this case those pertaining to request/repair algorithms in multicast, in order to provide some intuition to people implementing actual systems and to help guide their choices. In that vein, I would be interested in knowing more about any work that really directly builds off of or uses their idea of a framework for scalable, reliable multicast and learn in exactly what way this particular paper was useful. On an unrelated note, it also very seriously annoyed me that the graphs in this paper lacked axis or data labels of any kind.