Hi,

it might be that doing memory pageouts --- done to increase the vnode
cache --- hurt the io,
and breaking the clustering by that.
does the same happen if there is plenty of free mem?

lars



Chuck Silvers wrote:

> hi,
>
> yea, there is some performance work remaining to be done for softdeps
> in the UBC world.  (and other areas too, but this is one of the more
> noticable ones.)  I haven't had a chance to investigate yet,
> but it's on my list of things to do.
>
> -Chuck
>
> On Wed, Jan 03, 2001 at 02:46:57AM +0900, IWAMOTO Toshihiro wrote:
> > On the other day, I've noticed rather significant drop in tarball
> > extraction spped with UBC.
> >
> > Here are some figures. Tests are done on a softdep ffs with a 433MHz
> > Celeron machine which has 256MB of RAM.
> >
> > With an 1.5K kernel with 50MB of bufcache (without UBC):
> >     > tar zxf /usr/pkgsrc/distfiles/emacs-20.7.tar.gz
> >     4.438u 3.049s 0:19.26 38.7%     0+0k 424+8202io 58pf+0w
> >     > time tar zxf /usr/pkgsrc/distfiles/emacs-20.6.tar.gz
> >     4.565u 2.961s 0:14.27 52.6%     0+0k 381+8248io 2pf+0w
> >     > rm -rf emacs-20.[67]
> >     > time tar zxf /usr/pkgsrc/distfiles/emacs-20.6.tar.gz
> >     4.610u 3.061s 0:10.46 73.3%     0+0k 95+8191io 0pf+0w
> >     > time tar zxf /usr/pkgsrc/distfiles/emacs-20.7.tar.gz
> >     4.567u 2.993s 0:13.52 55.8%     0+0k 361+8214io 0pf+0w
> >     > rm -rf emacs-20.[67]
> >     > time tar zxf /usr/pkgsrc/distfiles/emacs-20.7.tar.gz
> >     4.541u 2.794s 0:12.60 58.1%     0+0k 92+8206io 0pf+0w
> >
> > With an 1.5N kernel with 40MB of bufcache (with UBC):
> >     > time tar zxf /usr/pkgsrc/distfiles/emacs-20.7.tar.gz
> >     4.459u 5.165s 0:43.43 22.1%     0+0k 141+385io 34pf+0w
> >     > time tar zxf /usr/pkgsrc/distfiles/emacs-20.6.tar.gz
> >     4.517u 5.087s 0:46.41 20.6%     0+0k 109+396io 1pf+0w
> >     > rm -rf emacs-20.[67]
> >     > time tar zxf /usr/pkgsrc/distfiles/emacs-20.7.tar.gz
> >     4.607u 5.132s 0:26.44 36.8%     0+0k 2+358io 0pf+0w
> >     > time tar zxf /usr/pkgsrc/distfiles/emacs-20.6.tar.gz
> >     4.506u 5.277s 0:45.34 21.5%     0+0k 2+407io 0pf+0w
> >
> >
> > If you run iostat while extracting tarballs, you might see the likely
> > cause of speed drop.  Without UBC, about 40k bytes are written per
> > transfer.  With UBC, writes are done in smaller chunks (7~9kB per
> > transfer).
> >
> > I'm not sure how much this speed drop affects normal usage.  Also I'm
> > not sure if this can be easily fixed.  I'm posting this message here
> > in the hope that someone takes look at this issue...
> >
> > --
> > IWAMOTO Toshihiro