Eric Dumazet wrote:

(2 weeks ago, Corey mentioned a 10x increase on UDP throughput on a 
16-way machine,
that sounds promising)

Just to be clear, that was 10x with preempt RT, which converts rwlocks 
into PI mutexes.  So 16 processors going for the same lock is pretty ugly.

Under heavy loads there is also a writer starvation problem, I believe 
in non-RT.  You can't actually create or destroy a UDP socket when the 
load is high because there's always a reader holding the lock.  RCU also 
solves that problem.

-corey