If there was an increase in the number of bytes charged to the socket
for each skb, without a corresponding increase in TCP receive buffer
size, that would explain a systematic decrease in receiver window that
would cause some paths, like perhaps these, to become
receiver-window-limited (which is consistent with the symptom of
stable bandwidth for a given path).

Can you please try increasing the default TCP receiver window to see
if this helps; try this as root on the 3.2 machine:

sysctl net.ipv4.tcp_rmem="4096 897664 897664"

Basically making the middle number as big as the third number should
bump up the default receiver window size.

neal