Re: Performance drop in Linux 2.4.19-XFS 1.2 ?

[Seth Mos <knuffie@xxxxxxxxx>]

At 14:17 30-7-2003 -0600, Craig Tierney wrote:
> On Wed, Jul 30, 2003 at 09:43:50PM +0200, Seth Mos wrote:
> > At 13:24 30-7-2003 -0600, Craig Tierney wrote:
> >
> > >Sorry if this is a repeat, I haven't seen the whole
> > >thread.  Did you try increasing the scsi max readahead?
> >
> > I found them already since new something like this existed in proc :-)
> >
> > >   echo "511" > /proc/sys/vm/max-readahead
> > >    echo "127" > /proc/sys/vm/min-readahead
> >
> > I used 512 and 128 but it works!
>
> Great.  I used 2^n-1 becuase that is how they
> were set initially (31).

The default changed between the release but the penalty on devices that can 
do decent fast IO is dramatic.
In my case it halves the read performance compared to what it used to be.

2.4.18-27
/proc/sys/vm/min-readahead 3
/proc/sys/vm/max-readahead 127

2.4.20-18
/proc/sys/vm/min-readahead 3
/proc/sys/vm/max-readahead 31

I think we want to change this, if your devices can pass the 100MB/sec read 
barrier this will probably affect you.

My suggested default would be 31 for min-readahead and 127 for 
max-readahead. I tested a bit and it gives decent results.
Anything over 127 is not really used. Below 127 it starts hurting the 
performance in large sequential IO.

The disks I have here are significantly fast that a min-readahead smaller 
then 31 just make them seek a lot. There is not much difference in latency 
to fetch this much information as it does for 3.

Cheers



--
Seth
It might just be your lucky day, if you only knew.