kernel history of ide performance

kernel history of ide performance

[kiu]

Hi ide team,

i often heard "uh, since kernel 2.6.bla the ide perfomance is soo bad".
Because of that (and the joy of writing bash scripts noone else will understand
*g*) i wrote an automated test script to get comparable measurements.

The script compiles 2.6.0-2.6.11.6 (yes "" | make oldconfig), boots each kernel,
runs hdparm/bonnie++ and create some pngs with gnuplot.

I ran the script on a Intel 82371AB/EB/MB PIIX4/PIICeleron 333/IC35L040AVER07
machine and you can find the results here:

    http://www.kiu.weite-welt.com/misc/idetest

I am quite busy at the moment and hopefully have time next month to clean up the
scripts for a release (are you interested?) and do further testing...
-- 
kiu

Re: kernel history of ide performance

[Bartlomiej Zolnierkiewicz]

Hi,

On Apr 5, 2005 10:46 AM, kiu <kiu@gmx.net> wrote:
> Hi ide team,
> 
> i often heard "uh, since kernel 2.6.bla the ide perfomance is soo bad".
> Because of that (and the joy of writing bash scripts noone else will understand
> *g*) i wrote an automated test script to get comparable measurements.

Please note that hdparm results can be highly influenced 
by changes in the Virtual Memory management.

One of my TODO items is tool for testing raw IDE performance,
so we know where to put blame for degraded results... :)
[ but it should wait until all IDE driver is taskfile based ]

> The script compiles 2.6.0-2.6.11.6 (yes "" | make oldconfig), boots each kernel,
> runs hdparm/bonnie++ and create some pngs with gnuplot.
> 
> I ran the script on a Intel 82371AB/EB/MB PIIX4/PIICeleron 333/IC35L040AVER07
> machine and you can find the results here:
> 
>     http://www.kiu.weite-welt.com/misc/idetest

Good job, thanks for doing this.

> I am quite busy at the moment and hopefully have time next month to clean up the
> scripts for a release (are you interested?) and do further testing...

It would be great to test at least every -rc kernel so eventual
regressions can be discovered (and fixed) before final release.

Bartlomiej

Re: kernel history of ide performance

[Oliver.Korpilla]

kiu wrote:
> Hi ide team,
> 
> i often heard "uh, since kernel 2.6.bla the ide perfomance is soo bad".
> Because of that (and the joy of writing bash scripts noone else will understand
> *g*) i wrote an automated test script to get comparable measurements.
> 
> The script compiles 2.6.0-2.6.11.6 (yes "" | make oldconfig), boots each kernel,
> runs hdparm/bonnie++ and create some pngs with gnuplot.
> 
> I ran the script on a Intel 82371AB/EB/MB PIIX4/PIICeleron 333/IC35L040AVER07
> machine and you can find the results here:
> 
>     http://www.kiu.weite-welt.com/misc/idetest
> 
> I am quite busy at the moment and hopefully have time next month to clean up the
> scripts for a release (are you interested?) and do further testing...

This looks like a mixed bag: some regression, some clear improvements, 
mostly the same. Nice work!

Does this portray the IDE situation or more the situation of the PIIX 
chipset - I guess the code isn't changed as much since it should be 
rather mature?

With kind regards,
Oliver Korpilla

Re: kernel history of ide performance

[Bartlomiej Zolnierkiewicz]

On Apr 5, 2005 11:02 AM, Oliver.Korpilla@gmx.de <Oliver.Korpilla@gmx.de> wrote:
> kiu wrote:
> > Hi ide team,
> >
> > i often heard "uh, since kernel 2.6.bla the ide perfomance is soo bad".
> > Because of that (and the joy of writing bash scripts noone else will understand
> > *g*) i wrote an automated test script to get comparable measurements.
> >
> > The script compiles 2.6.0-2.6.11.6 (yes "" | make oldconfig), boots each kernel,
> > runs hdparm/bonnie++ and create some pngs with gnuplot.
> >
> > I ran the script on a Intel 82371AB/EB/MB PIIX4/PIICeleron 333/IC35L040AVER07
> > machine and you can find the results here:
> >
> >     http://www.kiu.weite-welt.com/misc/idetest
> >
> > I am quite busy at the moment and hopefully have time next month to clean up the
> > scripts for a release (are you interested?) and do further testing...
> 
> This looks like a mixed bag: some regression, some clear improvements,
> mostly the same. Nice work!
> 
> Does this portray the IDE situation or more the situation of the PIIX
> chipset - I guess the code isn't changed as much since it should be
> rather mature?

Probably the situation of VM code...

Having scripts at hand would be useful to double check that 
testing procedure itself is OK.

Bartlomiej

Re: kernel history of ide performance

[kiu]

Quoting Bartlomiej Zolnierkiewicz <bzolnier@gmail.com>:

> On Apr 5, 2005 11:02 AM, Oliver.Korpilla@gmx.de <Oliver.Korpilla@gmx.de>

> > Does this portray the IDE situation or more the situation of the PIIX
> > chipset - I guess the code isn't changed as much since it should be
> > rather mature?

> Having scripts at hand would be useful to double check that
> testing procedure itself is OK.

You asked for it (ptest.tar.gz on the website) dont blame me for the chaos :)
-- 
kiu

Re: kernel history of ide performance

[Mark Lord]

Note that hdparm has a "--direct" flag to perform O_DIRECT I/O,
which can sometimes give different information than the default method.

But the one weakness with 'hdparm -t' in either case,
is that it performs a single READ at a time, rather than
queuing up a pipeline full of READs.  So it should always
slightly under report achievable throughput.

Cheers