Date: Wed, 4 Nov 2009 13:49:54 +0100 From: Mel Flynn <mel.flynn+fbsd.hackers@mailing.thruhere.net> To: freebsd-hackers@freebsd.org, rea-fbsd@codelabs.ru Subject: Re: Issue with grep -i (on i386 only?) Message-ID: <200911041349.54943.mel.flynn%2Bfbsd.hackers@mailing.thruhere.net> In-Reply-To: <T7DtLCOP0cwiekv/7ybsxY3l5dQ@7ANLw7WpNQUEViOFvqmcIRbmcl4> References: <200911032122.28905.mel.flynn%2Bfbsd.hackers@mailing.thruhere.net> <T7DtLCOP0cwiekv/7ybsxY3l5dQ@7ANLw7WpNQUEViOFvqmcIRbmcl4>
next in thread | previous in thread | raw e-mail | index | archive | help
On Wednesday 04 November 2009 04:05:44 Eygene Ryabinkin wrote:
> Mel, good day.
>
> Tue, Nov 03, 2009 at 09:22:28PM +0100, Mel Flynn wrote:
> > So on the laptop I modified the testscript as it is attached now and
> > while there is still a significant delay, the wallclock time is less
> > then half, when the expression is rewritten with the same meaning:
> > =>>> 16777216
> > =>>> fgrep
> > 0.04 real 0.03 user 0.00 sys
> > 0.05 real 0.03 user 0.01 sys
> > 0.02 real 0.00 user 0.00 sys
> > =>>> pcregrep
> > 0.26 real 0.21 user 0.02 sys
> > 0.26 real 0.22 user 0.02 sys
> > 0.44 real 0.35 user 0.01 sys
> > =>>> grep
> > 0.04 real 0.04 user 0.00 sys
> > 4.45 real 4.15 user 0.01 sys
> > 2.00 real 1.81 user 0.00 sys <-- [fF][Oo][Oo]
>
> Just did a quick test on the 8.0-RC2/i386 with very old Athlon processor:
> -----
> =>>> 16777216
> =>>> fgrep
> 0,09 real 0,04 user 0,05 sys
> 0,18 real 0,06 user 0,03 sys
> 0,05 real 0,01 user 0,04 sys
> =>>> pcregrep
> 0,47 real 0,29 user 0,07 sys
> 0,52 real 0,33 user 0,07 sys
> 0,77 real 0,45 user 0,03 sys
> =>>> grep
> 0,09 real 0,08 user 0,01 sys
> 0,10 real 0,04 user 0,05 sys
> 0,23 real 0,12 user 0,03 sys
> -----
> Pattern for the plain 'grep' is stable: first and second variants always
> give the same time within a 0.01 second variation and the last variant
> gives 2x slowdown.
>
> I tried sizes up to the 64M -- the pattern stays. The same stuff for
> the amd64, so in my case I don't see the difference in behaviour. So,
> maybe, the problem isn't 32 vs 64 but lies somewhere else.
Well, just ruled out the last commonality: The i386 machines tested all had
MAXPHYS to 1M, except the one I just tried:
=>>> 16777216
=>>> fgrep
0.04 real 0.03 user 0.00 sys
0.04 real 0.03 user 0.00 sys
0.02 real 0.00 user 0.01 sys
=>>> grep
0.04 real 0.02 user 0.02 sys
3.70 real 3.56 user 0.00 sys
1.91 real 1.83 user 0.02 sys
Using env MALLOC_OPTIONS= also has no impact at all (just in case defaults
aren't that). Since fgrep is fast and basically seeds the cache for grep, I'm
ruling out disks/io reads. In fact, /tmp on this laptop is memory disk (one
reason I couldn't go up to 64M :)). I honestly can't figure out what my 'local
problem' could be or your optimization.
Thanks for the fix ups. One more below sig.
--
Mel
--- grep-test.sh.orig 2009-11-04 03:17:05.000000000 -0900
+++ grep-test.sh 2009-11-04 03:29:55.000000000 -0900
@@ -34,6 +34,10 @@
;;
esac
+ if [ ! -f ${TMPFILE} ]; then
+ # signalled
+ exit 0;
+ fi
jot -r -c ${b} a z |rs -g 0 20 > ${TMPFILE}
echo "=>>> ${b}"
for prog in fgrep ${PCREGREP} ${BSDGREP} grep ; do
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?200911041349.54943.mel.flynn%2Bfbsd.hackers>