Strange raid behaviour

[Chris Eborn]

On 12/15/05, Gilboa Davara <gilboada at netvision.net.il> wrote:
> On Thu, 2005-12-15 at 17:07 +0000, Chris Eborn wrote:
> > Does the output from the following  commands seem odd to people?
> >
> > [root at nucoda root]# hdparm -t /dev/md0
> >
> > /dev/md0:
> >  Timing buffered disk reads:  774 MB in  3.01 seconds = 257.52 MB/sec
> > [root at nucoda root]# mount /dev/md0 /array
> > [root at nucoda root]# hdparm -t /dev/md0
> >
> > /dev/md0:
> >  Timing buffered disk reads:  416 MB in  3.00 seconds = 138.55 MB/sec
> > [root at nucoda root]# umount /array
> > [root at nucoda root]# hdparm -t /dev/md0
> >
> > /dev/md0:
> >  Timing buffered disk reads:  776 MB in  3.01 seconds = 257.42 MB/sec
> > [root at nucoda root]#
> >
> >
> > Tha array is made of 10 scsi disks - each of which will read at around
> > 60 MB/sec and all seems fine with the raid
> > device until the fs is mounted and it appears to halve the trasnfer
> > rate! There is no activity in the filesystem. At the moment this
> > is with FC2, though I have tried with FC3 and numerous kernels.
> > Tweaking around with block sizes and thing does change the speed of
> > the raw /dev/md0 -  but the speed always drops when I mount the
> > filesystem. Can anybody explain this behaviour?
> > Maybe I am missing something obvious, but this is baffling me - and
> > what is worse is that it all worked until I upgraded the system. I
> > have tried to go back, even put a new system disk in and started from
> > scratch, but cannot get a decent read speed from the array.
> >
> > Chris
> >
>
> I'm seeing the same behavior on my 4x36GB MD5 setup.
> Mounted FS: ~105-110MB/s.
> Unmount FS: ~125MB/s.
>
> Can you check the read-ahead settings (blockdev --getra /dev/md0)
> before /and/ after fs mount?
>
> Gilboa
>

Hi There,
Well I am still suffering with this - I have installed Fedora Core 2
with kernel 2.6.14 and the raid is still going half speed.
I have spent quite a while playing with the readahead values (blockdev
--setra) on both the individual disks and the raid device (/dev/md0)
and they do not seem to change anything, even with extreme values
(like "blockdev --setra 0 /dev/sd[abcdefghi]"). Leads me to think that
readahead has been disabled elsewhere.

Chris
ps - Unfortunately this machine needs to be able to show 24 12 meg
files a second (that's because film runs at 24fps) - so come the new
year it may well have to fall into the windows empire :-(