I've started collecting various data (including oprofile) using our
DBT-2 (OLTP) workload with lvm2 on linux 2.6.2 and 2.6.3 on ia32 and
ia64 platforms:
http://developer.osdl.org/markw/lvm2/
So far I've only varied the stripe width with lvm, from 8 KB to 512 KB,
for PostgreSQL that is using 8 KB sized blocks with ext2. It appears
that a stripe width of 16 KB through 128KB on the ia64 system gives the
best throughput for the DBT-2 workload on a volume that should be doing
mostly sequential writes.
I'm going to run through more tests varying the block size that
PostgreSQL uses, but I wanted to share what I had so far in case there
were other suggestions or recommendations.
--
Mark Wong - - [email protected]
Open Source Development Lab Inc - A non-profit corporation
12725 SW Millikan Way - Suite 400 - Beaverton, OR 97005
(503) 626-2455 x 32 (office)
(503) 626-2436 (fax)
http://developer.osdl.org/markw/
In article <[email protected]>,
<[email protected]> wrote:
>I've started collecting various data (including oprofile) using our
>DBT-2 (OLTP) workload with lvm2 on linux 2.6.2 and 2.6.3 on ia32 and
>ia64 platforms:
> http://developer.osdl.org/markw/lvm2/
For write-heavy loads, you might want to see if 2.6.4-rc2-mm1 makes
any difference. It fixes a case where both pdflush and a writing
process itself would queue writes to an LVM device, resulting in
less than optimal I/O ordering for some cases. The fix will probably
be in 2.6.4 proper as well.
Mike.
On Mon, 08 Mar 2004, [email protected] wrote:
> I've started collecting various data (including oprofile) using our
> DBT-2 (OLTP) workload with lvm2 on linux 2.6.2 and 2.6.3 on ia32 and
> ia64 platforms:
> http://developer.osdl.org/markw/lvm2/
>
> So far I've only varied the stripe width with lvm, from 8 KB to 512 KB,
> for PostgreSQL that is using 8 KB sized blocks with ext2. It appears
Does ext2 write 8 KByte sized blocks atomically on IA32?
Or is this no requirement for PostgreSQL consistency?
--
Matthias Andree
Encrypt your mail: my GnuPG key ID is 0x052E7D95
[email protected] wrote:
> I've started collecting various data (including oprofile) using our
> DBT-2 (OLTP) workload with lvm2 on linux 2.6.2 and 2.6.3 on ia32 and
> ia64 platforms:
> http://developer.osdl.org/markw/lvm2/
>
> So far I've only varied the stripe width with lvm, from 8 KB to 512 KB,
> for PostgreSQL that is using 8 KB sized blocks with ext2. It appears
> that a stripe width of 16 KB through 128KB on the ia64 system gives the
> best throughput for the DBT-2 workload on a volume that should be doing
> mostly sequential writes.
>
> I'm going to run through more tests varying the block size that
> PostgreSQL uses, but I wanted to share what I had so far in case there
> were other suggestions or recommendations.
>
Here's one thought: look at the i/o rates on individual drives using
each stripe size. You *might* see that one size does far fewer seeks
than others, which is a secondary thing to optimize after throughput IMHO.
If you don't have a tool for this I can send you the latest diorate
which does stuff like this, io rate perdrive or per partition, something
I occasionally find revealing.
-bill
On Thu, Mar 11, 2004 at 05:12:35PM -0500, Bill Davidsen wrote:
> [email protected] wrote:
> > I've started collecting various data (including oprofile) using our
> > DBT-2 (OLTP) workload with lvm2 on linux 2.6.2 and 2.6.3 on ia32 and
> > ia64 platforms:
> > http://developer.osdl.org/markw/lvm2/
> >
> > So far I've only varied the stripe width with lvm, from 8 KB to 512 KB,
> > for PostgreSQL that is using 8 KB sized blocks with ext2. It appears
> > that a stripe width of 16 KB through 128KB on the ia64 system gives the
> > best throughput for the DBT-2 workload on a volume that should be doing
> > mostly sequential writes.
> >
> > I'm going to run through more tests varying the block size that
> > PostgreSQL uses, but I wanted to share what I had so far in case there
> > were other suggestions or recommendations.
> >
> Here's one thought: look at the i/o rates on individual drives using
> each stripe size. You *might* see that one size does far fewer seeks
> than others, which is a secondary thing to optimize after throughput IMHO.
>
> If you don't have a tool for this I can send you the latest diorate
> which does stuff like this, io rate perdrive or per partition, something
> I occasionally find revealing.
Yeah, please do send me a copy. I'd be interested to see what that might
turn up. I've just been using iostat -x so far.
Thanks,
Mark
Is this generally available to all? Where can it be had?
TIA
On Thu, 2004-03-11 at 14:25, Mark Wong wrote:
> On Thu, Mar 11, 2004 at 05:12:35PM -0500, Bill Davidsen wrote:
> > [email protected] wrote:
> > > I've started collecting various data (including oprofile) using our
> > > DBT-2 (OLTP) workload with lvm2 on linux 2.6.2 and 2.6.3 on ia32 and
> > > ia64 platforms:
> > > http://developer.osdl.org/markw/lvm2/
> > >
> > > So far I've only varied the stripe width with lvm, from 8 KB to 512 KB,
> > > for PostgreSQL that is using 8 KB sized blocks with ext2. It appears
> > > that a stripe width of 16 KB through 128KB on the ia64 system gives the
> > > best throughput for the DBT-2 workload on a volume that should be doing
> > > mostly sequential writes.
> > >
> > > I'm going to run through more tests varying the block size that
> > > PostgreSQL uses, but I wanted to share what I had so far in case there
> > > were other suggestions or recommendations.
> > >
> > Here's one thought: look at the i/o rates on individual drives using
> > each stripe size. You *might* see that one size does far fewer seeks
> > than others, which is a secondary thing to optimize after throughput IMHO.
> >
> > If you don't have a tool for this I can send you the latest diorate
> > which does stuff like this, io rate perdrive or per partition, something
> > I occasionally find revealing.
>
> Yeah, please do send me a copy. I'd be interested to see what that might
> turn up. I've just been using iostat -x so far.
>
> Thanks,
> Mark
>
> _______________________________________________
> linux-lvm mailing list
> [email protected]
> https://www.redhat.com/mailman/listinfo/linux-lvm
> read the LVM HOW-TO at http://tldp.org/HOWTO/LVM-HOWTO/
Chris Croswhite wrote:
> Is this generally available to all? Where can it be had?
I stuck it on the page I use to avoid beating my fractional T1 to death.
http://pages.prodigy.net/davidsen/diorate.pl
Enjoy, any misfeatures please feed back to me.
--
bill davidsen <[email protected]>
CTO TMR Associates, Inc
Doing interesting things with small computers since 1979
In article <[email protected]>,
Mark Wong <[email protected]> wrote:
| On Thu, Mar 11, 2004 at 05:12:35PM -0500, Bill Davidsen wrote:
| > Here's one thought: look at the i/o rates on individual drives using
| > each stripe size. You *might* see that one size does far fewer seeks
| > than others, which is a secondary thing to optimize after throughput IMHO.
| >
| > If you don't have a tool for this I can send you the latest diorate
| > which does stuff like this, io rate perdrive or per partition, something
| > I occasionally find revealing.
|
| Yeah, please do send me a copy. I'd be interested to see what that might
| turn up. I've just been using iostat -x so far.
Okay, I posted the pointer a few days ago to LKML, did you get a chance
to try it? And if so, did it tell you anything?
--
bill davidsen <[email protected]>
CTO, TMR Associates, Inc
Doing interesting things with little computers since 1979.
On 16 Mar, bill davidsen wrote:
> In article <[email protected]>,
> Mark Wong <[email protected]> wrote:
> | On Thu, Mar 11, 2004 at 05:12:35PM -0500, Bill Davidsen wrote:
>
> | > Here's one thought: look at the i/o rates on individual drives using
> | > each stripe size. You *might* see that one size does far fewer seeks
> | > than others, which is a secondary thing to optimize after throughput IMHO.
> | >
> | > If you don't have a tool for this I can send you the latest diorate
> | > which does stuff like this, io rate perdrive or per partition, something
> | > I occasionally find revealing.
> |
> | Yeah, please do send me a copy. I'd be interested to see what that might
> | turn up. I've just been using iostat -x so far.
>
> Okay, I posted the pointer a few days ago to LKML, did you get a chance
> to try it? And if so, did it tell you anything?
I've grabbed it but I haven't had a chance to try it yet. Hopefully
later today or tomorrow.
Mark