David S. Miller a écrit :
Loopback is the wrong thing to optimize. We have several efficient
mechanisms for local system communications. If you don't necessarily
need socket semantics, use pipes (named or unnamed), else you can use
AF_UNIX sockets.
Van Jacobson shows a 6X improvement for real over the wire
communications, and you're arguing over a sub-percentile improvement
over loopback. There is a major difference.
Patches exactly like your's have been proposed before by the SGI folks
in the past, and I'm arguing against it using with the same positions
I held at that time.
What I want to point out is that kernel is *full* of hot points that are no
real reasons to exist/persist. Yes, they are better ways to exchange data
between two processes than using AF_INET sockets.
I sent for example a patch about the central rwlock_t that iptables/x_tables
is using. rwlock_t are traps for programmers that think read_lock() is
scalable, while it's not if used by several cpus.
http://lists.netfilter.org/pipermail/netfilter-devel/2006-January/023175.html
I got no answer from netfilter team, while this patch is in use on real
production servers. The CPUS have more cycles available for user level processing.
Linux took 5 years to get out of BKL (Big Kernel Lock), so I think things can
evolve step by step. In a few months I will use RCU for the x_tables problem
if nobody else did the job before me :)
-
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to [EMAIL PROTECTED]
More majordomo info at http://vger.kernel.org/majordomo-info.html
