Re: Removing softdep

[Gary Thorpe <gathorpe79%yahoo.com@localhost>]

--- On Mon, 6/9/08, Simon Burge <simonb%NetBSD.org@localhost> wrote:

> From: Simon Burge <simonb%NetBSD.org@localhost>
> Subject: Re: Removing softdep
> To: tech-kern%netbsd.org@localhost
> Received: Monday, June 9, 2008, 11:38 AM
> I've seen a few queries about speed of journaling vs
> softdep.
> 
> Here's a couple of sets of figures, both a few months
> old.  First, an
> old favourite - extracting pkgsrc.tar.gz:
> 
> # mount /dev/ld0e /mnt
> # tar xpfz ~/pkgsrc.pax.gz   1.489u 12.201s 18:29.87  1.2% 
>  887+544016io
> # rm -rf pkgsrc              0.115u  3.609s  9:46.81  0.6%
> 22870+228470io
> 
> # mount -o log /dev/ld0e /mnt
> # tar xpfz ~/pkgsrc.pax.gz   1.296u 10.531s  0:37.78 31.2% 
>  884+ 36157io
> # sync                       0.000u  2.593s  0:17.08 15.1% 
>    0+     7io
> # rm -rf pkgsrc              0.075u  3.415s  0:14.70 23.6%
> 21962+ 22898io
> # sync                       0.000u  0.004s  0:00.21  0.0% 
>    3+   370io
> 
> # mount -o softdep /dev/ld0e /mnt
> # tar xpfz ~/pkgsrc.pax.gz   1.555u 10.015s  0:33.00 35.0% 
>  886+ 22968io
> # sync                       0.000u  4.415s  0:12.42 35.5% 
>    3+  4358io
> # rm -rf pkgsrc              0.084u  1.387s  0:15.32  9.5%
> 21926+ 22850io
> # sync                       0.000u  1.221s  0:01.26 96.8% 
>    3+     2io
> 
> # mount -o async /dev/ld0e /mnt
> # tar xpfz ~/pkgsrc.pax.gz   1.426u  9.273s  0:20.66 51.7% 
>   41+ 22906io
> # sync                       0.000u  3.970s  0:08.04 49.3% 
>    9+     3io
> # rm -rf pkgsrc              0.125u  2.401s  0:12.29 20.5%
> 22051+ 22869io
> # umount /mnt
> 
> Async is fastest, and softdep is a little faster than
> journaling for
> extract and a little slower for removing.  All three are
> _much_ faster
> than normal ffs.
> 
> Now for the one workload I know of where journaling does
> seem to lose
> quite badly - repeated fsync() on files.  CVS does this for
> each file it
> touches and was how I discovered this "problem". 
> These are times for
> doing two rtags on a large repo, with /tmp and cvs lock
> dirs both on
> tmpfs:
> 
> cvsroot normal:
>    71.603u 115.288s 1:17:40.49 4.0%  0+0k 76624+1633298io
> 435012pf+0w
>    72.598u 115.665s 1:21:14.85 3.8%  0+0k 75899+1634869io
> 434696pf+0w
> 
> cvsroot softdep:
>    71.266u 112.522s   37:35.69 8.1%  0+0k 80044+ 818830io
> 434998pf+0w
>    71.840u 111.369s   40:19.31 7.5%  0+0k 78683+ 818657io
> 434804pf+0w
> 
> cvsroot log:
>    71.808u 133.269s 6:12:53.42 0.9%  0+0k 78479+3373981io
> 434963pf+0w
>    72.070u 132.084s 5:56:33.73 0.9%  0+0k 72534+3215163io
> 435128pf+0w
> 
> A really bad result for journaling there...  This also
> scaled down as
> well, just tagging a directory with a dozen files was
> measurably much
> slower with journaling.
> 
> Cheers,
> Simon.

Just a FYI: Sun benchmarked their UFS logging implementation against softdeps 
(I think) and the overall results where similar in that the performance wasa 
about the same with one edging the other in some categories (I think their 
implementation of logging did slightly better in their tests).


      __________________________________________________________________
Ask a question on any topic and get answers from real people. Go to Yahoo! 
Answers and share what you know at http://ca.answers.yahoo.com