On Tue, 7 Oct 1997, Tim Bass wrote:
> Bora> I don't know how one can implement this without a major tweak at the
> Bora> upper layers or letting network drivers snoop into
> Bora> device drivers but it would certainly be nice.
>
> Setting a congestion bit in TCP vs. a corruption bit?
>
> From a pure engineering perspective, please explain how this would be
> accomplished. On the surface, it seems like a major divergence
> from the TCP philosophy to attempt to have end-to-end systems
> distinguish between dropped packets due to congestion (buffer overflow)
> and corruption (noise).
>
> Is there a credible IETF paper on this?
>
> Tim
>
The idea of letting routers setting a congestion bit in times of
incipient congestion was described in a paper by Sally Floyd, "TCP and
Explicit Congestion Notification" published in the October 1994 issue of
CCR. It is also available by ftp as
ftp://ftp.ee.lbl.gov/papers/tcp_ecn.4.ps.Z
At the ipngwg meeting at the Munich IETF, Sally presented how you would do
this by using two bits in the IP header, a "ECN capable" bit and a
"Congestion Experienced" bit. The minutes mentioned that a draft should be
out a couple of week after the meeting but I haven't seen any yet.
Explicit Congestion Notification (ECN) is meant to be used together with
RED capable routers which may detect incipient congestion by monitoring
the average queue size.
Instead of dropping the packet when incipient congestion is detected
the router may set the "Congestion Experienced" bit in the header of a IP
packet, if it has the ECN-capable bit set.
By using a bit in the IP header instead of the TCP header, the ECN scheme
works for any transport protocol which implements congestion control, not
just TCP.
A TCP receiver that receives a packet with the "Congestion Experienced"
bit set, signals the TCP sender by adding a ECN-notify TCP option to the
acknowledgment. The TCP sender then reacts by halving its congestion
window, but only once per window of data.
The assumption that a lost packet indicates congestion remains though.
So to be able to detect corruption related losses a Explicit *Loss*
Notification (ELN) is needed like Hari Balakrishnan described in anonther
mail in this thread.
-- Christofer
--
Christofer Kanljung
Ericsson Mobile Data Design
[email protected]
This archive was generated by hypermail 2b29 : Mon Feb 14 2000 - 16:14:30 EST