From: Tyler Robison (trobison@cs.washington.edu)
Date: Tue Oct 12 2004 - 22:45:55 PDT
This paper argues that performing various functions at a low
level, below TCP, isn't a good idea; ideas such as reliable
data delivery and encryption should be handled by the application itself.
Many reasons are provided, one of the main points being the end to end
argument in which a function needs to have the high level knowledge of the
application itself to understand enough to successfully perform the
function, and other important ones being that performance gains for using
low level versions may not turn out to be gains in the long run, and that
there are various flaws in trusting every point of connection between the
sender and receiver with many tasks (since low level functions will be
carried out at every hop through the networks). In terms of encryption,
implementing it on a low level would require that more parties have access
to the keys, and that the data is more vulnerable for longer periods of
time.
The arguments presented here seem fairly persuasive, and a number
of good points are brought up, but everything is fairly abstract, and
there is a conspicuous absence of experimental data. They do give an
example of how low level checksum checking at each hop caused a rather
serious problem on a network at MIT, but while it helps express the
problem, a single case is hardly a proof. Again, the arguments given make
sense, but some data is needed to be more convincing, especially in the
section on performance; they talk about how low level implementations may
promise good performance and not deliver, but does that turn out to be the
case in real networks, and if so, how frequently?
At the same time, their points do bring to the foreground
particular problems that aren't entirely obvious, and their notion that
some functions must be done at a higher level in order to be done
correctly does strongly indicate that low level solutions won't always
work. Overall, the issues involved are important, and excellent points
are raised, but a less abstract argument is needed.
This archive was generated by hypermail 2.1.6 : Tue Oct 12 2004 - 22:45:56 PDT