From: Pravin Bhat (pravinb@u.washington.edu)
Date: Wed Nov 10 2004 - 07:53:34 PST
Paper summary: The paper proposes a new multicast architecture, Protocol Independent
Multicast (PIM), that unlike existing alternatives performs well in face of sparsely
distributed group members across wide areas allowing the technology to potentially
support multicast messaging over the entire internet.
Paper strength:
The paper clearly explains the motivations behind the need for a new multicast
architecture. The current flooding based techniques or the techniques that rely on
reconstructing the network topology (SPT) at each router cannot be efficiently
scaled to work in large networks. However these techniques work well for multicast
sessions that involve large fractions of the leaf nodes in a network. PIM tries to
strike a balance between SPT based unicast messaging and large group multicast messaging
by dynamically trying to group regions with dense distributions of group members
into a shared routing tree minimizing extraneous transmissions and at the same
time sparsely populated regions are served using a SPT based tree to minimize
flooding in these regions.
Limitations and Areas for improvement:
The authors claim that one of the major motivations for the deployment of
PIM over Core Based Trees (CBT) is that in practice the paths used by a single
shared tree tends to be up to 1.4 times longer than the paths that might be used
if a source specific SPT were to be constructed. However PIM cannot avoid such
path inflation anyway considering that its goal is to support multicasting for
wide areas spanning several ISPs/domains. Setting up all that state in intra-domain
routers to ensure multicast packets inside a domain are routed along optimal routes
might not be of much help when potentially one early-exit during inter-domain
routing could override any performance gains from using PIM.
Along the same lines the authors claim that using CBT causes greater traffic
concentrations in certain links. However the use of preset Rendezvous Points
and designated routers in PIM can cause similar stress in links connected to
these high traffic points. Concentration of traffic does not provide a strong
motivation for using PIM over CBT since this phenomenon is also common with
popular servers and several technologies(CDN, rotating DNS entries) exist to
solve this issue. It's easy to imagine generating several multicast shared trees
for each domain which are assigned in a round-robin fashion when different groups
want to start a multicast session thus distributing the load evenly over the
entire domain.
Also its not clear how a receiver can know when to initiate a shortest-path
distribution without knowing ahead of time how many hosts from its current
domain have already joined the multicast session. At the same time authors
do not really explain how the first hop routers decide when to switch
between SPT trees and shared trees.
I was surprised to see that paper didnt provide an empirical comparision of PIM
and any of other multicast routing techniques. I wasnt even sure if they had
a working system at the time of writing the paper as they provide no numbers
on the runtime performance of PIM. Overall I felt the paper could have done a
better job of explaining the PIM architecture. At times various sections in
the paper were hard to follow and didnt make a lot of sense.
Relevance and Future work:
Any significant work on multicasting protocols is bound to be relevant in terms
of providing efficient services for the types of applications that are gaining
popularity over the internet. Media broadcasts, teleconferencing, services that
involve flash crowds can all use an efficient multicast technology to significantly
reduce the amount of redundant traffic in the internet. The last paper we read
suggested that P2P traffic accounts for majority of the traffic over the internet
today and will continue to gain prominence in the years to come. One can see how
the combination of multicast protocols and tornado codes could potentially
minimize the amount of bandwidth that is wasted from having numerous users download
the same file across P2P networks.
For future work, the authors need to deploy PIM over several domains and compare
its performance to existing technologies.
This archive was generated by hypermail 2.1.6 : Wed Nov 10 2004 - 07:53:38 PST