On Wed, 1 Sep 2004 17:06:41 +0530
Srivatsa Vaddagiri <vatsa@xxxxxxxxxx> wrote:
> On Tue, Aug 31, 2004 at 03:54:20PM +0200, Andi Kleen wrote:
> > I bet also when you just do rdtsc timing for the TCP receive
> > path the cycle numbers will be way down (excluding the copy).
>
> I got cycle numbers for the lookup routine (with CONFIG_PREEMPT turned off).
> They were taken on a 900MHz 8way Intel P3 SMP box. The results are as below:
>
>
> -------------------------------------------------------------------------------
> | 2.6.8.1 | 2.6.8.1 + my patch
> -------------------------------------------------------------------------------
> Average cycles | |
> spent in | |
> __tcp_v4_lookup_established | 2970.65 | 668.227
> | (~3.3 micro-seconds) | (~0.74 microseconds)
> -------------------------------------------------------------------------------
>
> This repesents improvement by a factor of 77.5%!
And yet none of your benchmarks show noticable
improvements, which means that this micro-measurement
is totally unimportant in the grand scheme of things
as far as we know.
I'm not adding in a patch that merely provides some
micro-measurement improvement that someone can do a
shamans dance over. :) If we're going to add this
new level of complexity to the TCP code we need to
see some real usage performance improvement, not just
something that shows up when we put a microscope on
a single function.
|