Re: 2.5.67-mm4

[Nick Piggin <piggin@cyberone.com.au>]

Toon van der Pas wrote:

>On Fri, Apr 18, 2003 at 01:45:36AM -0700, Andrew Morton wrote:
>
>>ftp://ftp.kernel.org/pub/linux/kernel/people/akpm/patches/2.5/2.5.67/2.5.67-mm4/
>>
>>. A bunch of anticipatory scheduler patches.
>>
>>  For the first time ever, AS is working well with both IDE and SCSI
>>  under all the usual tests.
>>
>>  It works just fine on SCSI with zero TCQ tags, and with four TCQ tags. 
>>  At eight tags, read-vs-write performace is starting to measurably drop off.
>>  At 32 tags it is about 2000x slower than at zero or four tags.
>>
>>  My recommendation, as always, is to disable SCSI TCQ completely.  If you
>>  really must, set it to four tags.
>>
>
>What about drivers that bypass de SCSI layer?
>I administer a server with a Mylex RAID controller (DAC960)...
>
I'd say they will show the same behaviour. As far as we can tell
its an interaction between badly behaving TCQ disks (those which
allow lots of writes to bypass a read), and the anticipatory
scheduler. We will be working on fixing this so those inclined
can use huge numbers of outstanding tags without things going
too haywire.