It could be L2 cache-coloring effects as well.  Try to keep the working
set size smaller than the L2 cache size of the cpu you are on.

When/if using netperf, it will send from and recv to a "ring" of buffers one 
larger than the reported (as opposed to requested) socket buffer size divided by 
the send/recv buffer size.  Controlled via the -W global command-line option.

rick jones