I wasn't thinking of the actual *allocation* overhead; I was thinking
of the working set of (unswappable) pages needed to play with a bunch
of existing "struct sock"s. Perhaps this isn't as much of an issue as
I first assumed, but I see all that extra time spent in "free_pages"
and "_get_free_pages", and I imagine the kernel working twice as hard
to find memory for everything else simply because its pool of "struct
sock"s is twice as big as it used to be.
Kevin <buhr@stat.wisc.edu>