On Mon, 29 Nov 1999, Jakob �stergaard wrote:
> On Mon, Nov 29, 1999 at 02:58:42AM -0500, Remo Strotkamp wrote:
> > Hi there,
> >
> > first of all thanx for all the replies concerning my previous mail.
> > second: hi jakob and here are the single mode bonnie results...
>
> > expect the raid1 beeing lots faster,
> > as reads are spread in parallel over lots more of disks!!!!!! But it
> > seems to be alot slower....
>
<stuff deleted>
> Contrary to what the HOWTO says (ouch!) RAID-1 doesn't (yet) distribute
> reads over the disks. RAID-5 does (has to).
>
> Mika Kuoppala is working on a patch that makes RAID-1 distribute reads over the
> disks. Hopefully in the not-too-distant future, you will see better
> performance from RAID-1 in many cases.
>
Actually RAID-1 does distribute the reads over the disks if they
are non-sequential. Sequential reads are read only from single disk.
(bonnie :)
The way that RAID-1 code does distribute them is simple. After every
read the disk is changed. From what i have studied this seems
to end up in unnecessary seeking and thus degrades performance.
My patch tries to minimize seeking by bookkeeping disk head positions
for every disk in array. They are not physical positions, rather a linear
logical positions, but this approach seems to make some difference
regardless. Both read throughput and seek times
are better. The patch can be found http://www.iki.fi/miku/raid1
Thus far (due to lack of brave testers like you!) i have only
one reference system. You can read more about how it performs with
and without this patch at: http://www.iki.fi/miku/raid1/tests
Jakob said that he might be able to give it a try too, so
with his permission i will add the results in above page.
> It's pretty hard to make RAID-1 run faster in a bonnie test, as bonnie works
> with one file only. But in a production environment, many files are usually
> accessed simultaneously, and it should be possible for RAID-1 to scale there.
>
Straight to the point here. This is what i have tried to achive.
Benchmarks usually give only limited picture (like bonnie in raid1 case).
I made a simple concurrent io tester for developing this patch
(www.iki.fi/miku/tiotest) but nevertheless benchmarks with real io
intensive application would reveal us the whole truth :)
-- Mika <[EMAIL PROTECTED]>
