From: Susumu Harada (harada@cs.washington.edu)
Date: Sun Oct 10 2004 - 17:23:09 PDT
"A Digital Fountain Approach to Reliable Distribution of Bulk Data"
J. Byers et al.
The paper proposes an alternate solution for reliable multicasting of
digital data over generic network medias using a framework termed "digital
fountain" and an algorithm for encoding the data called the Tornado codes.
The term "digital fountain" reflects the nature of the protocol in that
a client interested in retrieving the data from the server can join in at
any time during the broadcast and tap in for duration roughly equivalent
to the period in which the data would have been delivered had the
connection been point to point.
The paper presents a well organized description of the problem domain,
outlining the desired characteristics of an ideal digital fountain,
currently known protocol (Reed-Solomon codes with interleaved codes), and
a thorough performance comparison with the proposed Tornado code. The
empirical support for the performance data are very impressive and
meticulous, with significant improvement over the alternative. It is
great that they not only included simulation results with multiple
experimental conditions, but also results based on actual trace data
(although the authors acknowledge that it was limited in scope).
I found the approach of the randomized graph-based coding of the Tornado
code to be very intuitive and its simplicity appealing. The layered
multicast approach was also interesting, although it seems that they did
not get as much mileage from it as I expected.
One point of concern I had was regarding the way in which the Tornado code
was based on random generation of the graph structure. They state that
"for all 10,000 trials the same graph was used... one might achieve better
performance by testing various random graphs for performance before
settling on one". What is not clear is how dependent the performance
result of a particular graph structure is on the structure of the data
itself. Since in real application the data that is being multicast would
be varying from one case to the next, the dependence on sample-and-pick
for determining the suitable graph structure for each data set does not
seem too realistic.
I would also like to see how the protocol could possibly be improved by
incorporating some degree of client feedback to increase the efficiency of
the overall network usage by the multicasting host. Although the appeal
of the proposed protocol is that it works without the need for the clients
to respond to the server, it would be interesting to see if minimal
interaction can be introduced to even further the efficiency of the
algorithm.
This archive was generated by hypermail 2.1.6 : Sun Oct 10 2004 - 17:23:10 PDT