Experiences from the Operational Deployment of an Overlay-based Internet Broadcasting System
*********************************************************************************************


Since the initial proposals for Overlay/Application/End System
Multicast, there have been several proposed designs, backed by
simulation experiments and experiments on small-scale wide-area
Internet test-beds, that have indicated the promise of the
architecture. However, there has been relatively little effort on
building complete operational systems used by real users in
representative Internet environments. Such an operational system
would both validate the Overlay Multicast architecture, as well
as provide traces and data that can guide future research.

In this poster, I will present the efforts of our research group in
building a completely operational ESM system that is being used by
*real* users in *real* applications. This system has already been used
to broadcast real conferences and workshops (Sigcomm 2002, Sigcomm
2003, AID 2003), Distinguished Lectures at Carnegie Mellon, and a
broadcast to the Slashdot community. It has already been used by over
three thousand people, spread over various continents, in commercial,
educational and home settings, behind NATs and firewalls, and with
various Operating Systems including Linux, MAC and Windows. To our
knowledge, this is the most comprehensive real deployment of a research
Overlay Multicast system to date.

I will present key implementation/engineering decisions in building such
an operational ESM system, and challenges involved in setting up such a
research vehicle. I will then present important insights and
observations that we have been learning from this research vehicle
which is driving our future research.

I am applying the intuition gained from real experience to my research
in two ways. First, I am using it model real world phenomena that we
have found important in practice - but are often neglected by current
simulation studies - such as node heterogeneity (DSL Vs. Ethernet),
presence of NATs and firewalls, bandwidth path heterogeneity, Internet
dynamics, and practical issues in adapting to metrics like bandwidth.
Second, I am conducting simulation studies of several Overlay
Multicast designs using these new models to help analyze these designs.
Such studies can expose important issues in the designs, given that
they model realistic scenarios that are generally not considered
in simulation experiments today.


As an example of the insight the real-world experience has helped
derive, I will present the "scalable resource location" problem.  In
Overlay Multicast protocols, a host in response to group and network
dynamics may need to select a new parent (resource). We have observed
that typical Internet environments are resource scarce, with a large
fraction of DSL and NATs, factors not usually considered in current
simulation experiments. This scarcity of resources makes it difficult
to locate a new parent, and the protocol must have explicit features
for efficiently locating new parents in resource-scarce environments -
a topic that has not been paid attention by many designs today. I will
present results that demonstrate the importance of this problem for
various families of Overlay Multicast protocols, and solutions that
can apply across various designs.