Re: Linux File system and Oracle setup

Fabrizio apparently said,on my timestamp of 12/06/2005 5:27 PM:

> In 2.6 the ioscheduler changed and lvtune is now deprecated.

Thanks for confirming, that's what I thought I'd seen.

> You have to mess with sysfs now:
> /sys/block/<device>/queue/iosched

Oh no, NOT another universe of tuning rules... ;D

>
> Attention! Database servers, especially those using "TCQ" disks should
> investigate performance with the 'deadline' IO scheduler. Any system
> with high
> disk performance requirements should do so, in fact."

TCQ? That is more relevant at controller level, isn't it? Why would the anticipatory IO scheduler bother with it when it'd be a lot better handled by things like SANs? Got a feeling someone in Loonicks-land is trying to be too smart again.

Why don't they just concentrate in providing efficient aio and let the complexities of disk scheduling be handled in much better places? There *IS* a reason why mainframes eat all other environments in IO throughput: they concentrate only on what's important, instead of flights of fancy...

> And again:
>
> "For some workloads the anticipatory scheduler is around 10% slower
> than deadline. Most notably, database workloads which seek all over the
> disk performing reads and synchronous writes. Database folks will likely
> want to boot with elevator=deadline to get that last bit of
> performance back."

Holy Macka! Just what we needed: nobody knows for sure what happens when! I wonder which workload doesn't seek "all over the disk"? :)

Thanks for the useful info. I feel another round of reading internal stuff coming up. What do you reckon: kerneltrap?

-- 
Cheers
Nuno Souto
in sunny Sydney, Australia
wizofoz2k_at_yahoo.com.au.nospam