Prev: select_catalog_role does not let me see data dictionary . why?
Next: Collection
From: DG problem on 25 Jun 2010 01:45

> I almost always separate out the disks for online logs from the rest
> of the system.  For commit happy systems I really want to have average
> log file sync times at 1 ms or under if I can.

So even though Oracle recommends putting everything on the same
stripe, you still separate the log files? Have you noticed any
performance gains from doing this from your testing?

The thing is that being able to write out log files across an 8 disk
stripe means much faster write times if the IO isn't being chugged
down by other processes.
From: John Hurley on 25 Jun 2010 06:37
DG:

> So even though Oracle recommends putting everything on the same
> stripe, you still separate the log files? Have you noticed any
> performance gains from doing this from your testing?
>
> The thing is that being able to write out log files across an 8 disk
> stripe means much faster write times if the IO isn't being chugged
> down by other processes.

You have to test in your own environment and measure what different
setups work better or not so much. As I noted my systems are OLTP
databases with some very commit happy applications ( commit too
much ... like on every row often ).

The one size fits all recommendations are generic ones that may or may
not make sense. Like I noted I get an average of less than 1 ms for
log file sync ... that is the wait event I tune for on the online log
files.

I am also using for my most important prod systems RAID 1 for the LUNs
for the log files and RAID 1+0 for the rest of the database portions
( again using ASM for both ). Any competent storage array buffers the
write activity in cache.

For test databases and our datawarehouse systems we are using RAID
5 ... not my prod OLTP systems.

Yes you have to do a lot of work to try different setups and write
different things to simulate different kind of workloads. Custom
designed tests take a lot of time but things that go after your
important tables in ways that ( somewhat ) simulate your real
applications give better and ( hopefully ) more reliable performance
numbers than generic testing tools that are available.

From: Frank van Bortel on 29 Jun 2010 15:26
On 06/23/2010 04:07 PM, Mladen Gogala wrote:
> On Wed, 23 Jun 2010 11:00:11 +0200, Frank van Bortel wrote:
>
>> Make it all SAME: Stripe and Mirror Everything.
>
> You're in marketing? SAME == RAID 1+0.

No; and I prefer RAID 0+1.
Not in storage, I seem to recall striping was RAID-0, mirroring
was RAID-1. Same would be RAID-0+1, then, eh?
>
>> Don't know about stripe sizes -
>
> That was the OP's question.
>
>> what do you think to gain by playing with these?
>
> Speed.
>

Still, no one provided any hard numbers.

Everybody seems convinced about gains in playing
with these figures, but the question remains:
- is 1MB stripe any different from what the
vendor recommended?
- if so, what gains in speed were measured
ordered by stripe size?


--

Regards,

Frank van Bortel
From: joel garry on 29 Jun 2010 16:17
On Jun 29, 12:26 pm, Frank van Bortel <fbor...(a)home.nl> wrote:
> On 06/23/2010 04:07 PM, Mladen Gogala wrote:
>
> > On Wed, 23 Jun 2010 11:00:11 +0200, Frank van Bortel wrote:
>
> >> Make it all SAME: Stripe and Mirror Everything.
>
> > You're in marketing? SAME == RAID 1+0.
>
> No; and I prefer RAID 0+1.
> Not in storage, I seem to recall striping was RAID-0, mirroring
> was RAID-1. Same would be RAID-0+1, then, eh?

No, see what the faq says, and that 1+0 is better (for fault
tolerance): http://www.orafaq.com/wiki/RAID

http://jonathanlewis.wordpress.com/2007/02/05/go-faster-stripes/ has
some good thoughts.

>
>
>
> >> Don't know about stripe sizes -
>
> > That was the OP's question.
>
> >> what do you think to gain by playing with these?
>
> > Speed.
>
> Still, no one provided any hard numbers.
>
> Everybody seems convinced about gains in playing
> with these figures, but the question remains:
> - is 1MB stripe any different from what the
>    vendor recommended?
> - if so, what gains in speed were measured
>    ordered by stripe size?

That's the gist: measure what the real load will be doing. I'd agree
with John about online logs, but most of my experience over the last
decade has been with RAID-5, so I didn't say anything 'cause I really
haven't looked. And yes, I'm a baarf member.

This one seems to say go coarse on the stripes:
http://www.hds.com/assets/pdf/best-practices-library-deploying-oracle-11g-rac-with-hds.pdf
I haven't looked at it too closely to see if it makes sense, and
haven't checked whether it applies to the OP hardware. "The default
stripe depth for an ASM group is 1MB, which is too low to make
efficient use of the underlying RAID group stripe size. This should be
set to 8MB at ASM group creation time by specifying the AU_SIZE
attribute..."

jg
--
@home.com is bogus.
"I believe it’s horribly bad in SQL Server, but it’s doubly horribly
bad in Oracle." - Tom Kyte on triggers

From: Frank van Bortel on 3 Jul 2010 09:41
On 06/29/2010 10:17 PM, joel garry wrote:
> On Jun 29, 12:26 pm, Frank van Bortel<fbor...(a)home.nl> wrote:
>> On 06/23/2010 04:07 PM, Mladen Gogala wrote:
>>
>>> On Wed, 23 Jun 2010 11:00:11 +0200, Frank van Bortel wrote:
>>
>>>> Make it all SAME: Stripe and Mirror Everything.
>>
>>> You're in marketing? SAME == RAID 1+0.
>>
>> No; and I prefer RAID 0+1.
>> Not in storage, I seem to recall striping was RAID-0, mirroring
>> was RAID-1. Same would be RAID-0+1, then, eh?
>
> No, see what the faq says, and that 1+0 is better (for fault
> tolerance): http://www.orafaq.com/wiki/RAID

1+0 it is.
>
> http://jonathanlewis.wordpress.com/2007/02/05/go-faster-stripes/ has
> some good thoughts.
>
>>
>>
>>
>>>> Don't know about stripe sizes -
>>
>>> That was the OP's question.
>>
>>>> what do you think to gain by playing with these?
>>
>>> Speed.
>>
>> Still, no one provided any hard numbers.
>>
>> Everybody seems convinced about gains in playing
>> with these figures, but the question remains:
>> - is 1MB stripe any different from what the
>> vendor recommended?
>> - if so, what gains in speed were measured
>> ordered by stripe size?
>
> That's the gist: measure what the real load will be doing. I'd agree
> with John about online logs, but most of my experience over the last
> decade has been with RAID-5, so I didn't say anything 'cause I really
> haven't looked. And yes, I'm a baarf member.
>
> This one seems to say go coarse on the stripes:
> http://www.hds.com/assets/pdf/best-practices-library-deploying-oracle-11g-rac-with-hds.pdf
> I haven't looked at it too closely to see if it makes sense, and
> haven't checked whether it applies to the OP hardware. "The default
> stripe depth for an ASM group is 1MB, which is too low to make
> efficient use of the underlying RAID group stripe size. This should be
> set to 8MB at ASM group creation time by specifying the AU_SIZE
> attribute..."
>
> jg
> --
> @home.com is bogus.
> "I believe it�s horribly bad in SQL Server, but it�s doubly horribly
> bad in Oracle." - Tom Kyte on triggers
>

Will check it out - any hard numbers?

--

Regards,

Frank van Bortel
First  |  Prev  | 
Pages: 1 2
Prev: select_catalog_role does not let me see data dictionary . why?
Next: Collection