From: Susumu Harada (harada@cs.washington.edu)
Date: Mon Nov 15 2004 - 02:59:11 PST
"Internet Indirection Infrastructure"
I. Stoica et al.
This paper proposes an network overlay abstraction dubbed the Internet
Indirection Infrastructure (i3) to provide a foundation for supporting
mobility, multicasting and anycasting by decoupling the senders from
receivers. They call this the rendezvous-based communication abstraction,
where the sender and the receiver "rendezvous" at a midpoint of their
communication, which in this case is an i3 server (or set of servers)
which serve to translate between each end-hosts's IP addresses and their
corresponding unique i3 identifier. By having the i3 servers take care of
managing this mapping between identifiers and IP addresses, and by
allowing receivers to add or modify "triggers" (mapping between identifier
and the IP address), the abstraction allows for flexibility in enabling
multicasting and end-host mobility to be implemented on top of it.
The paper does a great job of addressing many key issues concerning the
scalability, security and efficiency of i3. They propose several
solutions for each of the main areas of concern, such as using multiple
private triggers for secure communication flow between end hosts and using
a challenge to verify request for addition of new triggers to prevent
spurious trigger inflation.
One of the several points at which I thought the paper could have done a
better job was in testing their system under a more realistic setting.
Their experiment involved only a few dozen i3 servers running on what
seemed to be a very small local network. Since their proposal involves
the use of this framework for supporting multicasting and mobility, I
would like to see how it actually performs in an Internet-scale network
with millions of hosts communicating to each other and attempting to
communicate using the trigger servers. I would like to know how many i3
servers will be necessary in order to support traffic of such scale. It
was also not clear how the receiver and the host goes about sharing
information on what identifier to use for a particular multicast session
for example.
Another concern I have is in their proposed solution for avoiding a single
point of failure by use of backup triggers or replicating triggers across
servers. Again, with the large scale of Internet traffic, I would like to
see empirical evidence to show how it can scale even with such redundancy.
I am particularly interested to know how well this infrastructure would
work for supporting mobile clients, and how it compares to typical Mobile
IP implementation in terms of performance and robustness. This issue will
become more important as the number of mobile agents on the Internet
continues to grow.
This archive was generated by hypermail 2.1.6 : Mon Nov 15 2004 - 02:59:12 PST