>One topic that I think we need to keep in mind as we talk about how smart or
>stupid TCP is about link-layer corruption is that we need to have solutions
>that work in non-trivial network topologies.
Yes, but this doesn't mean that we should reject out of hand any
special-case enhancements. Certain special cases, such as a TCP in a
single computer directly attached to a slow radio link, are extremely
common and are worth a special look.
>The alternative, using a Performance-Enhancing Proxy to "split" the TCP
>connection at the border of the wireless and wireline networks, does allow
>both wireless-interface TCPs to get link-level notifications directly - but
>we understand a lot less about the impact of split-TCP operation on
>end-to-end throughput than we do about end-to-end TCP.
As a long-time disciple of the End-To-End Principle, these things
really make me uncomfortable. They introduce state, which makes
robustness harder to achieve, and they make it hard to use various
security mechanisms on an end-to-end basis -- all serious, if not
fatal, problems with the WAP (Wireless Application Protocol, aka Why
Another Protocol) project, IMHO.
I'd want to exhaust all the possibilities of link-specific performance
enhancements, any appropriate generic enhancements to TCP, and even
application-specific tuning of TCP before resorting to something that
does as much violence to the Internet model as a transport proxy.
Phil
This archive was generated by hypermail 2b29 : Mon Jan 28 2002 - 09:12:20 EST