A Reliable Multicast Framework for Light-weight Sessions and Application Level
Framing
Floyd, et al.

This paper presents a multicast analogue of TCP which relaxes some of the
constraints of previous proposals. 

The authors of this paper suggest that the existing generic reliable
multicast protocols are insufficient flexible for general use by applications.
While TCP works well for unicast communication, it doesn't extend well to
multicast, primarily due to scalability issues.  To address this, the authors
propose Scalable Reliable Multicast (SRM).  

The SRM framework assumes IP multicast, unique persistent names for data, and
the ability to separate the application's name space into 'pages'.  In SRM,
members periodically multicast session messages to maintain state.  For
reliable delivery, SRM has a loss recovery algorithm in which members that
have missed a packet multicast a repair request and members which have
received the packet multicast a repair response.

In the loss recover scheme, many requests and responses may turn out to be
unnecessary, so it works best when the request and response timer parameters
are tuned to avoid those unnecessary packets.  This path latencies for a given
network topology may be uniform or quite varied.  Since the tuning can have a
significant effect on the recover time, the authors outline an adaptive
algorithm which attempts to self-tune the mechanism.  Another optimization
explored by the authors is limiting the recovery communication to some subset
of the group so that traffic to the entire group can be avoided.

SRM makes many assumptions about the application.  Two assumptions which stand
out are the naming scheme for the data and that the data represent idempotent
operations.  I question how many WAN-level distributed systems obey these
assumptions.  The demonstrated application, wb (for White Board), seems like a
relatively limited case, and the authors don't investigate what existing
applications could use it.  

I see the benefit of SRM over n^2 TCP connections in terms of network
overhead, though I wonder how many WAN-level applications have IP multicast
available and need this protocol.  Outside of the SRM framework itself, I
thought the data on loss recovery was interesting, and I would expect some of
the principles to apply well to other distributed communication protocols.