Re: RAID 0+1 and recommended stripe size

Bricklen <BAnderson_at_nospamthanks.com> wrote in message news:<3E79E27D.68F67D33_at_nospamthanks.com>...
> Thanks Connor, for the confirmation, and to Joel as well for the SAME
> article.
> I will definitely try the suggestions I've received and benchmark for
> our system.
>
> btw, I notice in that SAME article that Joel has provided:
>
> http://otn.oracle.com/deploy/availability/pdf/oow2000_same.pdf
>
> that Juan Loaiza et al. suggests using 1 meg stripes, as well as a 1 meg
> db_file_multiblock_read_count, and OS IO size limits to 1 meg.
>
> I'm wondering if anyone has tried both methods and has an informed
> opinion on which is _generally_ more performant? Mind you, I'm not
> after some magical incantation like fast=true, just a push in the right
> direction to start off my testing/benchmarking.

The point of the same article is to avoid the question - which means, don't use same if you truly care about every little bit of performance. To paraphrase what they are saying, it is generally not cost-effective to put a lot of time into it, as opposed to believing their incantations. I think this is appropriate for small systems. Where it becomes inappropriate is highly debateable. As a generalization, SAME is not more performant - the whole point is that the marginal increase in performance isn't worth the time taken to do a more formal methodology. In that, I disagree with the paper, because I think their results are a byproduct of Oracle's bias towards cpu-binding, and may already be inapplicable to current hardware. But I agree for small systems because, while an argument can be made that a small system would benefit most from performance tuning (a technical argument I agree with), in practical terms there usually isn't a large staff to support such an effort. More likely it's one or two guys that have to make it work and can't justify dedicating their careers to tweaking this little system. It's simply cheaper to buy bigger hardware than more people. Sound familiar?

>
> My reasoning: I'm pressed for time on this issue so the option to try
> every possible way is out of the question.

And don't forget to ask your vendor what everyone else with similar configurations is doing (of course, they may all be doing it wrong :-P ).

Note that if it were up to me, the technical arguments would win every time. But that is not the way of the world.

>
> Thanks again everyone for the suggestions!
>
> Cheers,
>
> Bricklen
>
> Connor McDonald wrote:
> > The general consensus around the place is:
> >
> > block size * multiblock read * n
> >
> > where n = 1, 2, 3, ... etc
> >
> > Which 'n' is best? For *your* system it could be any of them, if you
> > can run some benchmarks you'll find the best for your system. It might
> > match the thoughts out there in the public forums, it might not...But
> > you will at least know thats its the best for *your* system
> >
> > hth
> > connor
> > --
> > =========================
> > Connor McDonald
> > http://www.oracledba.co.uk
> >
> > "Some days you're the pigeon, some days you're the statue"

jg

--
@home.com is bogus.
Don't miss the Fox Network's newest game show "Who Wants To Be A
Despotic Dictator?"