[cfgeeks] Combining hardware and software RAID on Solaris
Gil Young
gjyoung at cfl.rr.com
Mon Dec 11 12:37:33 EST 2006
Thanks all, doing a bit more research, and the practice has a name!
"They've gone to plaid!" Spaceballs, the movie (Oh I need to rent that
movie again.)
http://web.archive.org/web/20050426183754/http://sunsolve7.sun.com/kmsattachments/23101.plaid.pdf
Anyhow, I agree the KISS principle does have great merit. The devices
are unchanging, and what I plan on doing is configuring it how I would
normally with mirrors and raid 5 on all four, then striping a pair of
them in software, then running Bonnie++ or some other benchmark tool TBD
on both configs with differing file sizes and R/W characteristics. The
idea is to collect as much as possible data wise while not hindering
deployment too much. If I see, oh, 10% or greater gains for my
application i'll probably document it and run the more complex way.
I had done this once before for the previous incarnation of this
database, and I saw the performance peaks and valleys (and cliffs) at
the cache size of the RAID device, and the installed memory of the
system, it was pretty interesting. but this is the first time i'll have
enough RAID chassis to think about trying a mix of hardware and software
RAID.
Oh and i'm pushing it because I can. If I can see a 10% improvement, it
will also save time in export/restore windows and get me home earlier
over the next 7 years of the life of this hardware (when applicable),
and save the end users time as well, make the work force more
productive, then i see it as worth it for the few extra hours of test
and possible rebuild back to standard if it is not worth it.
Can't remember if I posted this here or another list, but it's something
like this:
"The story goes that the engineer working on the boot sequence for the
original Mac was working late one night when Steve Jobs wanders past and
asks how long the thing takes - the engineer is pretty happy that he's
gotten it down to around 30 seconds (or however long it was) and that's
probably good enough. Jobs then comments that they'll probably sell at
least a million of these things - and each one will probably be booted a
couple of times a day - and the machines will last maybe five years - so
if he can save just one second more from the bootup time - that's
equivelent to 113 years from the lives of Mac owners. So if you can save
just one more second - that's like saving someone's life."
Except for this the millions of boot sequences will be every time
someone does a transaction with the database (pr/po, hr, reports, etc..)
Liston Bias wrote:
> I have seen some gains from striping a set luns on different sets of
> redundant raid disks through testing... If you have hardware device
> dedicated to your server (i.e. Clariion dedicated to single server)
> that will very likely never change then it may be worth explorer.
> However, the gains were minimal.
>
> Nothing is my world is static, and the pains from configurating this
> kind of thing wrong are not worth the risk to me. It makes
> maintainance and response to change more time consuming. Are we
> pushing the envelope because we can or because the application
> requires it?
>
> I imagine the gains would improve as the things like write cache of
> your SAN device are decreased.
>
> Regards,
> Liston
>
>
> On Sun, 10 Dec 2006, Gil Young wrote:
>
>> I have four FC chassis coming in, each will have it's own dedicated
>> HBA and reside in/on a V890. This got me to thinking, has anyone
>> ever done a combo of hardware and software RAID for performance
>> gains? Were the gains appreciable?
>>
>> I'm thinking that it may be a good article/study, at the very least
>> tested with simple read/write benchmarks to see how it fares.
>>
>> _______________________________________________
>> cfgeeks mailing list
>> cfgeeks at mail.cfgeeks.org
>> http://mail.cfgeeks.org/mailman/listinfo/cfgeeks
>>
> _______________________________________________
> cfgeeks mailing list
> cfgeeks at mail.cfgeeks.org
> http://mail.cfgeeks.org/mailman/listinfo/cfgeeks
>
>
More information about the cfgeeks
mailing list