Re: ORACLE on Linux - IO bottleneck

Fabrizio Magni wrote:
> personally I have never played with the elvtune so I'm really interested
> in your result.
> Have you published something?

Alas, no time to publish anything. Got to get this thing running fast or else company can't charge clients! We get paid by the click: we lose click summaries, we can't charge, my bonus goes awol! :(

I'll see if I can put some text together in another month or so, the US guys are taking over the rest of the ops and I should be able to concentrate a bit more on doco for all the db stuff of late. Will check if AUSOUG is interested in it.

And I finally got a test box to figure out aio with cooked io in RHAS3 and 9i! There's gotta be a way of making it work good. And 10gr2 and RHAS4 to test as well: it's gonna be a good year!

> I tested the four 2.6 I/O schedulers (and their parameters) and the
> difference in performance are quite relevant.
> Benchmarks on the use of these schedulers are easy to spot out.
> I'm posting what was written on the linux kernel developer mailing list:
> http://www.cs.rice.edu/~ssiyer/r/antsched/shines.html
> http://www.cs.rice.edu/~ssiyer/r/antsched/antio.html
> and linux symposium (with a little bit of theory):
> http://www.linuxsymposium.org/proceedings/reprints/Reprint-Axboe-OLS2004.pdf
> http://www.linuxsymposium.org/proceedings/reprints/Reprint-Pratt-OLS2004.pdf

Thanks, good stuff. Yeah, looks like a major improvement. The only thing that worries me a bit is it seems to be totally oriented to native controllers or jbod architectures. We use mainly SANs with hyperchannel switches, which means I'll have to test this all over again!... Received on Sat Feb 11 2006 - 05:24:21 CST