[linux-lvm] md + lvm + reiser

[linux-lvm] md + lvm + reiser

[Chris Jantzen]

Excuse me if this has been covered already, though I did try to read
through the archives. I am running md raid5 (software) over several
drives and importing that into LVM and using reiserfs. It's worked fine
for some time, but now I want to start snapshots. I've learned about the
VFS lock patch, yes, but my real question is: why is the raid5 buffer
cache constantly thrashing back and forth? I assume this has to do with
a conflict between reiserfs's 4096 byte blocks and LVM's 512
SECTOR_SIZE. Or did I configure something wrong somewhere? Is it safe to
just change SECTOR_SIZE in lvm.h or will that mess up on-disk
structures? Thank you.

-- 
chris jantzen kb7rnl =->         __O
Insert witty comment here.     _`\<,_
http://www.maybe.net/         (*)/ (*)

Re: [linux-lvm] md + lvm + reiser

[Xavier Montagutelli]

On Sat, 7 Sep 2002, Chris Jantzen wrote:

> Excuse me if this has been covered already, though I did try to read
> through the archives. I am running md raid5 (software) over several
> drives and importing that into LVM and using reiserfs. It's worked fine
> for some time, but now I want to start snapshots. I've learned about the
> VFS lock patch, yes, but my real question is: why is the raid5 buffer
> cache constantly thrashing back and forth? I assume this has to do with
> a conflict between reiserfs's 4096 byte blocks and LVM's 512
> SECTOR_SIZE. Or did I configure something wrong somewhere? Is it safe to
> just change SECTOR_SIZE in lvm.h or will that mess up on-disk
> structures? Thank you.

I have exactly the same symptoms with the same configuration (md + lvm +
reiser) ; the raid5 cache size keeps changing when I start using
snapshots. Does someone have some tips to avoid/explain this ?


--
[ Xavier Montagutelli          `""'                                   ]
[ INSA Toulouse                0--0                                   ]
[               ---------oOOo---()---oOOo----------                   ]
[ Dpt de Math / Service Informatique  Tel: +33 5 61 55 93 38 / 67     ]
[ 135, Avenue de Rangueil             Fax: +33 5 61 55 93 20 / 60     ]
[ 31077 Toulouse Cedex 4              Mel: montagut@gmm.insa-tlse.fr  ]

Re: [linux-lvm] md + lvm + reiser

[Luca Berra]

On Sun, Sep 08, 2002 at 11:44:41PM +0200, Xavier Montagutelli wrote:
>On Sat, 7 Sep 2002, Chris Jantzen wrote:
>
>> VFS lock patch, yes, but my real question is: why is the raid5 buffer
>> cache constantly thrashing back and forth? I assume this has to do with
...
>I have exactly the same symptoms with the same configuration (md + lvm +
>reiser) ; the raid5 cache size keeps changing when I start using
>snapshots. Does someone have some tips to avoid/explain this ?

The r5 cache needs a fixed request size, Neil once explained me, but i
do not recall the exact details now, when the md layer receives a
request with different size, it has to be emptied and recreated with
the new size.
This is the meaning of the message, it is harmless, except for
performance, at the moment the best that can be done about it is
commenting out the message.

regards,
L.
-- 
Luca Berra -- bluca@comedia.it
        Communication Media & Services S.r.l.
 /"\
 \ /     ASCII RIBBON CAMPAIGN
  X        AGAINST HTML MAIL
 / \

Re: [linux-lvm] md + lvm + reiser

[Chris Jantzen]

On Mon, Sep 09, 2002 at 12:04:11AM +0200, Luca Berra wrote:
> On Sun, Sep 08, 2002 at 11:44:41PM +0200, Xavier Montagutelli wrote:
> >On Sat, 7 Sep 2002, Chris Jantzen wrote:
> >
> >>VFS lock patch, yes, but my real question is: why is the raid5 buffer
> >>cache constantly thrashing back and forth? I assume this has to do with
> ...
> >I have exactly the same symptoms with the same configuration (md + lvm +
> >reiser) ; the raid5 cache size keeps changing when I start using
> >snapshots. Does someone have some tips to avoid/explain this ?
> 
> The r5 cache needs a fixed request size, Neil once explained me, but i
> do not recall the exact details now, when the md layer receives a
> request with different size, it has to be emptied and recreated with
> the new size.
> This is the meaning of the message, it is harmless, except for
> performance, at the moment the best that can be done about it is
> commenting out the message.


Well, I know it's harmful and I know that's exactly why it happens. I
want my performance though! :-) I was wondering if any of the deep LVM
guru's could say whether SECTOR_SIZE could be safely bumped up. Thanks
for the input and I think I'll wait a few more days before making a hard
decision.

-- 
chris jantzen kb7rnl =->         __O
Insert witty comment here.     _`\<,_
http://www.maybe.net/         (*)/ (*)