From: Ethan Katz-Bassett (ethan@cs.washington.edu)
Date: Sun Oct 10 2004 - 19:44:22 PDT
In this paper, the authors use their Tornado codes to tackle the problem of
efficient mass distribution of large data over the Internet. The authors
present the digital fountain as the ideal solution to the problem. The
requirements of a digital fountain seem appropriate to the problem. Their
approach using Tornado codes approximates a digital fountain, and they
present experimental results comparing the Tornado codes to other schemes.
The approach bests earlier attempts by having more efficient encoding and
decoding (Reed-Solomon codes are MDS but computationally more complicated
than Tornado codes) and by scaling well to large files. I like that the
paper includes a description of an idealized solution as well as empirical
comparisons to other solutions; these contrasts frame the work well.
The paper does not address how much variations in the random graph
structures affect results. Since the efficiency depends on the particular
random graph used, the pretesting of various graphs to pick an efficient one
seems like a good idea. The area is not one I am familiar with, and I am
surprised that good solutions do not require interleaving. It was unclear
from the paper whether interleaving codes might outperform Tornado Z under
different conditions.
The problem is an important one. Techniques for retransmission that require
requests from the receiver, such as sliding-window, do not translate well to
multicast problems. However, widespread dissemination of large files
continues to become more important, and we need feasible, efficient
solutions. Throughout the paper, the authors mention various directions for
future work. Four seem most interesting to me: 1) In many conceivable
applications, mirrored sources seem likely; the authors have yet to address
mirroring effectively using Tornado codes; 2) Real-time multicast is an
important concern, and latency requirements change the characteristics of an
idealized solution; 3) The multicast layering subscription scheme seems to
necessitate "optimizing properties of the schedule further; 4) The authors
mention using Tornado codes to improve dispersity routing schemes.
This archive was generated by hypermail 2.1.6 : Sun Oct 10 2004 - 19:44:28 PDT