[e2e] TCP Local Area Normal behaviour? any references?

[Matt Mathis]

> Capture effects affect fairness, not throughput. In fact, by allowing a
> pair of stations to jump on the channel faster, capture effects enhance
> the throughput.

,,,and this is where theory and reality diverge.

During the 1GE roll-out, half duplex FastE was a major headache for people trying
to send large data sets.  Here is a pathological case:

A long path with GigE and OC-3 (or better) all the way from our data center to
the hub at the remote site.  The last span was likely to be a point to point
half duplex link from the hub to some users workstation.

During slow-start, our server and the long path deliver bursts to the far hub
which are at least 100 Mb/s.  Assuming there is enough buffering, nothing gets
dropped, but each burst captures the channel, locking out all ACKs until the
end of the burst.  If the last switch is properly buffered, TCP will progress
all the way into congestion avoidance in this mode:  A full window of data on
the forward path, followed by a full window of ACKs on the return path.  If you
set the windows and buffering properly, you get about half of the expected
performance.  If something goes wrong (like insufficient buffering), you get,
shall we say "interesting" behaviors and even worse performance.

Generalizing this situation: any link which exhibits any sort of
capture/lockout/modulation and is the dominant bottleneck in a path which is
both high delay an high capacity has the potential to be problematic.  If it is
the very first span the stack might be able to do something "hackish" to fix
it, but in all other cases there is nothing constructive for the stack to do
that does not hurt performance on healthy channels.

For the half duplex case packet traces are so bizarre that anybody who looked
at them immediately understood that something was seriously wrong.  For this
reasons most I2 connected sites are still rabid about eradicating both half
duplex and unmanaged hubs in peoples offices (the prime culprit these days).

Thanks,
--MM--
-------------------------------------------
Matt Mathis      http://www.psc.edu/~mathis
Work:412.268.3319    Home/Cell:412.654.7529
-------------------------------------------
Evil is defined by people who think they know
"The Truth" and use force to apply it to others.