Fwiw, I got very assymetic bahaviour between 2 i386/1.3.2
boxes, both on 100Mbit. One has an Intel 82558 chip, the other
a 3com905.
In one direction I got 7.x MByte/s, in the other some tens
of kByte. (I'm not sure which was which.)

This was not so in 1.2.x times, so I think some of the recent
TCP optimizations cause bad effects under certain load figures.
I'd try to switch the direction and play with the blocksizes.

best regards
Matthias