> It would also be hard to imagine that many link layers will adapt to
> TCP requirements. It would
> seem more of a managable problem to have TCP adapt to different link
> behaviors. Therefore, the
> suggestion below to have the document include the link layer
> characteristics of various communication
> links is a very good one.
TCP is an end-to-end protocol, designed to run over paths which can comprise
several vastly different links at the same time (e.g. it is very well
possible to have a "long thin" wireless access link, "short fat" wireline
networks and a "long fat" satellite link in the path your TCP/IP packets are
travelling along). It is probably hopeless to try and adapt TCP in a way
that it handles all of these vastly different link characteristics optimally
at the same time (this does of course not prohibit certain optimizations
which are beneficial for specific link types but don't decrease performance
for others).
Assuming this is true, two basic approaches are left: Either you split up
your TCP connection and optimize TCP on each segment of the path
individually (split-connection, spoofing, performance-enhancing proxies or
whatever you call it), or you adapt each individual link so that it fulfills
the assumptions made by TCP. The latter approach seems more sensible to me.
Of course there are certain "hard" characteristics of wireless links which
cannot be overcome by any protocol, but there are usually some degrees of
freedom for the designers of link and physical layers (e.g. tradeoff between
FEC and ARQ, channel coding vs. delay etc.).
Stephan Baucke
Ericsson Research
This archive was generated by hypermail 2b29 : Mon Jan 28 2002 - 09:12:19 EST