From: Danny Wyatt (danny@cs.washington.edu)
Date: Tue Nov 09 2004 - 21:09:26 PST
The PIM Architecture for Wide-Area Multicast Routing
Stephen Deering, Deborah Estrin, Dino Farinacci, Van Jacobson,
Ching-Gung Liu, Liming Wei
This paper presents a middle ground between two existing multicast
architectures. The first existing architecture, SPT (shortest path
tree), used the shortest routes between sender and receiver, but it
required periodically flooding the entire network with packets to
discover these paths. Clearly, that is not possible at the scale of
today's internet. The second existing architecture, CBT (core-based
tree), avoids flooding the entire network by establishing a core
rendezvous point for all traffic in a multicast group. The downside of
CBT is that this can drag traffic between some nodes to an out of the
way core that is not on the shortest tree spanning all of them.
The proposed PIM architecture merges these approaches by creating
rendezvous points similar in function to CBT cores. The PIM RP's manage
group and subscription info and initially forwards data between sender
and receiver. However, routers between receiver and sender can "break
off" form the tree defined by the RP's and establish shorter paths to
the source. This path-adaptation brings PIM closer to SPT.
At a very high-level, the protocol seems sensible enough---but they
provide no simulations of theory to support their claims. Additionally,
it is a very "engineered" approach that would require substantial
modifications to many, many routers on the internet. That there is
still no widespread multicast architecture in place today (nearly a
decade after this paper was written) suggests how likely such an
architecture is to succeed. Finally, I think I'm missing some detail
about pure SPT. As far as I know, calculating the minimum spanning tree
that encompasses a given set of nodes (a Steiner tree) is an NP-complete
problem. Is this what SPT is doing?
This archive was generated by hypermail 2.1.6 : Tue Nov 09 2004 - 21:09:29 PST