From: Ethan Phelps-Goodman (ethanpg@cs.washington.edu)
Date: Tue Nov 09 2004 - 17:46:11 PST
The PIM Architecture for Wide-Area Multicast Routing
Deering et. al.
This paper presents a new protocol for multicast routing. Existing protocols
of this time fell into three categories. The simplest, reverse-path
forwarding (RPF) just floods packets over the entire internet until they hit
a router that knows the location of one of the senders. This only scales in
the "dense" case, where you are communicating with a sizable fraction of the
internet. What they term shortest-path tree (SPT) protocols are analogous to
link-state routing. Each router maintains a full view of the network
topology and the position of the senders and receivers. This allows SPT to
build efficient distribution trees, but at the cost of maintaining a lot of
state. Finally, shared-tree schemes use a centralized router, which all
sources and receivers register with. This concentrates traffic at one point,
but greatly reduces shared state.
I got lost in the details of the protocol, which I don't think were well
explained. The overall idea, as I follow it, is that distribution trees
should start as shared-trees. As traffic flows, each router along the path
can break off from the shared-tree to a source specific SPT. The change is
incremental, allowing SPT to be used for high traffic applications, and
shared-tree for low traffic ones.
As far as I can tell, the scheme requires all intermediary routers to
understand the multicast protocol. Each router also has to maintain a
significant amount of state (which they cite as a main open problem.) They
also give no simulation data or theoretic analysis to justify their design
choices.
Ethan
This archive was generated by hypermail 2.1.6 : Tue Nov 09 2004 - 17:46:16 PST