Thanks, Eric, for the clarification.

I am still intrigued by these aspects of the issue:

Also I want to pay attention that limited speed stays constant - always
5.5 MB/s with straight line on traffic plotter as if was network bandwidth shaping.
With 3.1 kernels this never happens - speed can vary in some ranges
but at maximum it is about 11 MB/s, no limitations such as 5.5 MB/s.

...

But I can say that probably there is really 100Mb/s
with Full duplex because as I wrote in first mail
there is some servers which have almost the same
speed with 3.2.3 kernel as with 3.1.10.
They are mostly located geographically close to me
(maybe there are other servers, not close, but I not found such one).
Here example of such server ftp.linux.kiev.ua,
download speed 11,1 Megabytes/s with both kernels.

...

To me, these symptoms smell like a receiver window issue that is
capping bandwidth at some receive_window/RTT value:

1) when there is a problem bandwidth is a "straight line" instead of
typically variable TCP performance

2) when there is a problem, it is with geographically remote
(high-RTT) sites and not local sites

3) the problem only shows up when certain versions of the kernel act
as a receiver

I would be intrigued to see the following pieces of info for a slow
transfer (from a geographically distant server) with the problematic
3.2 kernel:

- tcpdump of first 300 or so packets to see SYN/SYNACK and RTT

- tcpdump of last 300 or so packets, to see the steady-state dynamics
  with receiver window and packet inter-arrival times

- summary of overall throughput wget sees

(I apologize if this is already in the thread somewhere... there are
now a lot of traces in the thread and it's hard to scan them all...
:-)

neal