>>> Tim Bass said:
> Hari> When these dupacks pass back through the base station,
> Hari> it sets the ELN bit in the TCP header, .......
>
> Okay! I agree that the appropriate place to set the flag
> is in the TCP header. Passing state information about the link
> (i.e. low BER, etc.) to the transport layer makes perfect sense.
Presumably you mean high BER? Actually, I'd rather not pass information
about BER per se, but only indicate the correct reason for segment loss, to
enable the right kind of recovery to happen.
> This state information is loosely-coupling the layers (worst
> case) and only marginally violates any end-to-end religious
> cannons.
Absolutely. The end-to-end religion, which I mostly believe in, permits
such "violations" as long as they only involve soft state and are a
performance enhancement. Another example of such a protocol very commonly
used today is good old VJ C-SLIP header compression, and it greatly
enhances performance!
> Which lends to the next question: How many competing TCP algorithms
> using FEC information are in the community today? Is there a
> short laundry list?
I don't think I understand this question. Could you please clarify it for
me? As far as I know, not too many TCP implementations use FEC. The only
transport protocols that I know of using FEC are some reliable multicast
protocols, in different stages of experimentation in the research community.
Are you talking about the potential interaction of FEC with TCP
retransmissions, or of the interactions between link-level ARQ schemes and
TCP? Or are you asking if there are different proposals to do things
similar to the ELN scheme I described before? If this is the question,
then I know of at least one other instance (where their algorithms and
environment were different, but the information needed was roughly the
same), by people at MITRE & Gemini (Durst, Miller & Travis). They had a
paper at Mobicom 96 on a protocol called SCPS-TP (which was mentioned
earlier on this thread, and which I'm sure many on this list are familiar
with).
Given our positive experience and theirs (in different networks), I think
it will be very useful to have an ELN bit (called "corruption experienced"
in SCPS-TP) in the TCP header for better performance in many situations.
-- Hari.
This archive was generated by hypermail 2b29 : Mon Feb 14 2000 - 16:14:30 EST