svn commit: r272906 - in head/sys: conf libkern netpfil/pf sys
Luigi Rizzo
rizzo at iet.unipi.it
Thu Oct 16 23:01:08 UTC 2014
On Wed, Oct 15, 2014 at 08:14:19AM +0400, Gleb Smirnoff wrote:
> On Tue, Oct 14, 2014 at 11:38:23PM -0400, Ed Maste wrote:
> E> On 14 October 2014 23:04, Gleb Smirnoff <glebius at freebsd.org> wrote:
> E> >
> E> > Look at results at the end of email. Guess what -j and -m mean. I
> E> > want either proving me doing the test wrong, or backing the change
> E> > out ASAP.
> E>
> E> It looks like there is indeed an error in the test:
> E>
> E> > h = murmur3_aligned_32((uint32_t *)&key[i],
> E> > sizeof(struct pf_state_key_cmp)/sizeof(uint32_t),
> E>
> E> murmur3's size argument is bytes, not uint32_ts, so the test is only
> E> hashing the first 1/4 of the keys.
>
> Thanks a lot, Ed. I rerun the tests. Results are fine. Distribution
> is equal (plots attached).
>
> Murmur is 9% +/- 1.5% faster. Sorry for noise.
is that 9% on the hash alone ?
If so it is surprising that you could see any change on pf's throughput.
In my very non-scientific tests with VALE, doing two hashes on
the MAC address would result in 17 Mpps vs 20 Mpps when no hashes were done.
The difference is about 8ns, which is totally in the noise at 2 Mpps or less.
cheers
luigi
> x jenkins-time
> + murmur-time
> +------------------------------------------------------------------------------+
> | ++ x |
> | ++ xx |
> |+ +++ xx |
> |+++++ + + xx x x x x x|
> | |_A_| |_____M______A_____________| |
> +------------------------------------------------------------------------------+
> N Min Max Median Avg Stddev
> x 12 8147243 8851519 8173132 8296753.8 235815.07
> + 15 7502767 7655925 7547376 7551089.5 39524.149
> Difference at 95.0% confidence
> -745664 +/- 127010
> -8.98742% +/- 1.53084%
> (Student's t, pooled s = 159194)
>
> --
> Totus tuus, Glebius.
More information about the svn-src-head
mailing list