FAQ Search Today's Posts Mark Forums Read
» Video Reviews

» Linux Archive

Linux-archive is a website aiming to archive linux email lists and to make them easily accessible for linux users/developers.


» Sponsor

» Partners

» Sponsor

Go Back   Linux Archive > Redhat > Fedora Development

 
 
LinkBack Thread Tools
 
Old 06-10-2008, 02:44 PM
Alan Cox
 
Default LVM negates benefits of jounaling filesystems?

On Tue, Jun 10, 2008 at 09:46:33AM -0400, Chuck Anderson wrote:
> filesystem metadata being written out to disk that jounaling
> filesystems rely on? If so, should we perhaps rethink the decision to
> use LVM by default on Fedora installs?

I would agree with that regardless (although LVM barrier handling if it is
broken does want fixing - I was under the impression it was ok). My boxes
are much faster without LVM so I always turn it off.

--
fedora-devel-list mailing list
fedora-devel-list@redhat.com
https://www.redhat.com/mailman/listinfo/fedora-devel-list
 
Old 06-10-2008, 02:47 PM
Alan Cox
 
Default LVM negates benefits of jounaling filesystems?

On Tue, Jun 10, 2008 at 09:05:34AM -0500, Eric Sandeen wrote:
> Barriers do have performance impact, which is probably historically why
> they're off by default on ext3. I'm not totally convinced that it's a

The cost is dropping rapidly with SATA devices. With a modern SATA disk and
NCQ we really really want them on - you've got 31 commands in flight and 16MB
or so of drive cache that will be written back in *arbitary* order. Thats
a recipe for disaster. A few years ago with one command at a time wind up
disks and at most 2MB it wasn't so bad. It really should be getting set on by
either our kernel or the installer in the SATA cases.

Alan

--
fedora-devel-list mailing list
fedora-devel-list@redhat.com
https://www.redhat.com/mailman/listinfo/fedora-devel-list
 
Old 06-10-2008, 02:49 PM
Eric Sandeen
 
Default LVM negates benefits of jounaling filesystems?

Alan Cox wrote:
> On Tue, Jun 10, 2008 at 09:46:33AM -0400, Chuck Anderson wrote:
>> filesystem metadata being written out to disk that jounaling
>> filesystems rely on? If so, should we perhaps rethink the decision to
>> use LVM by default on Fedora installs?
>
> I would agree with that regardless (although LVM barrier handling if it is
> broken does want fixing - I was under the impression it was ok). My boxes
> are much faster without LVM so I always turn it off.

LVM barriers aren't so much broken as simply un-implemented by design.

static int dm_request(struct request_queue *q, struct bio *bio)
{
...
/*
* There is no use in forwarding any barrier request since we can't
* guarantee it is (or can be) handled by the targets correctly.
*/
if (unlikely(bio_barrier(bio))) {
bio_endio(bio, -EOPNOTSUPP);
return 0;
}

... and somebody should probably measure the lvm overhead in general, I
suppose.

-Eric

--
fedora-devel-list mailing list
fedora-devel-list@redhat.com
https://www.redhat.com/mailman/listinfo/fedora-devel-list
 
Old 06-10-2008, 02:53 PM
Eric Sandeen
 
Default LVM negates benefits of jounaling filesystems?

Alan Cox wrote:
> On Tue, Jun 10, 2008 at 09:05:34AM -0500, Eric Sandeen wrote:
>> Barriers do have performance impact, which is probably historically why
>> they're off by default on ext3. I'm not totally convinced that it's a
>
> The cost is dropping rapidly with SATA devices. With a modern SATA disk and
> NCQ we really really want them on - you've got 31 commands in flight and 16MB
> or so of drive cache that will be written back in *arbitary* order. Thats
> a recipe for disaster. A few years ago with one command at a time wind up
> disks and at most 2MB it wasn't so bad. It really should be getting set on by
> either our kernel or the installer in the SATA cases.
>
> Alan
>

I think the problem is, barriers are really implemented as drive cache
flushes. So under certain workloads the performance really does hurt.
But if the alternative is a good chance of filesystem corruption on
power loss, remind me again why we run a journaling fileystem at all?

If ext3 doesn't get barriers on by default upstream then I would suggest
that yes, we should patch the kernel ourselves to do so, or set the
installer to create fstabs which specify it for filesystems that don't
have barriers on by default. ... after lvm stops filtering it out, anyway.

-Eric


--
fedora-devel-list mailing list
fedora-devel-list@redhat.com
https://www.redhat.com/mailman/listinfo/fedora-devel-list
 
Old 06-10-2008, 02:57 PM
Alan Cox
 
Default LVM negates benefits of jounaling filesystems?

On Tue, Jun 10, 2008 at 09:49:16AM -0500, Eric Sandeen wrote:
> if (unlikely(bio_barrier(bio))) {
> bio_endio(bio, -EOPNOTSUPP);
> return 0;
> }
>
> ... and somebody should probably measure the lvm overhead in general, I
> suppose.

About 30% for disk performance on my test box. Obviously that isn't all
some LVM code corner case and there are other problems in there to reach
that difference on dbench. That's FC8 and I believe some lvm fixes for
performance have occurred since I did the runs (FC8 no updates).

Alan

--
fedora-devel-list mailing list
fedora-devel-list@redhat.com
https://www.redhat.com/mailman/listinfo/fedora-devel-list
 
Old 06-10-2008, 03:56 PM
Bruno Wolff III
 
Default LVM negates benefits of jounaling filesystems?

On Tue, Jun 10, 2008 at 09:49:16 -0500,
Eric Sandeen <sandeen@redhat.com> wrote:
>
> LVM barriers aren't so much broken as simply un-implemented by design.
>
> static int dm_request(struct request_queue *q, struct bio *bio)
> {
> ...
> /*
> * There is no use in forwarding any barrier request since we can't
> * guarantee it is (or can be) handled by the targets correctly.
> */

That seems weird. I thought one of the reasons for having stacked block devices
is that you could pass on barrier requests.

I have seen comments about adding barrier support to linear block devices
that currently don't support it (e.g. dmcrypt) relatively recently.

On a somewhat related note there was a discussion about issues with barriers
on lmkl last February that suggested there are issues with sync on linux
if you have write cache enabled even if you are using barriers.

--
fedora-devel-list mailing list
fedora-devel-list@redhat.com
https://www.redhat.com/mailman/listinfo/fedora-devel-list
 
Old 06-10-2008, 04:02 PM
Eric Sandeen
 
Default LVM negates benefits of jounaling filesystems?

Bruno Wolff III wrote:
> On Tue, Jun 10, 2008 at 09:49:16 -0500,
> Eric Sandeen <sandeen@redhat.com> wrote:
>> LVM barriers aren't so much broken as simply un-implemented by design.
>>
>> static int dm_request(struct request_queue *q, struct bio *bio)
>> {
>> ...
>> /*
>> * There is no use in forwarding any barrier request since we can't
>> * guarantee it is (or can be) handled by the targets correctly.
>> */
>
> That seems weird. I thought one of the reasons for having stacked block devices
> is that you could pass on barrier requests.

Not really; in general it is trickier w/ more devices I think.

md for example passes barriers on raid1 but not 0 or 5 AFAIK...

> I have seen comments about adding barrier support to linear block devices
> that currently don't support it (e.g. dmcrypt) relatively recently.
>
> On a somewhat related note there was a discussion about issues with barriers
> on lmkl last February that suggested there are issues with sync on linux
> if you have write cache enabled even if you are using barriers.

Got a url?

Thanks,
-Eric

--
fedora-devel-list mailing list
fedora-devel-list@redhat.com
https://www.redhat.com/mailman/listinfo/fedora-devel-list
 
Old 06-10-2008, 04:21 PM
Eric Sandeen
 
Default LVM negates benefits of jounaling filesystems?

Alan Cox wrote:
> On Tue, Jun 10, 2008 at 09:49:16AM -0500, Eric Sandeen wrote:
>> if (unlikely(bio_barrier(bio))) {
>> bio_endio(bio, -EOPNOTSUPP);
>> return 0;
>> }
>>
>> ... and somebody should probably measure the lvm overhead in general, I
>> suppose.
>
> About 30% for disk performance on my test box. Obviously that isn't all
> some LVM code corner case and there are other problems in there to reach
> that difference on dbench. That's FC8 and I believe some lvm fixes for
> performance have occurred since I did the runs (FC8 no updates).
>
> Alan
>

quick test; fast hardware raid on fibrechannel, 4 cpu opteron, 8G
memory, 4-threaded dbench for 600s, 120s warmup, 500G ext3 fs, 2.6.26-rc2.

ext3 on /dev/sdc:
Throughput 679.495 MB/sec 4 procs

ext3 on /dev/mapper/myvg-lvol0 on /dev/sdc:
Throughput 665.599 MB/sec 4 procs

-Eric

--
fedora-devel-list mailing list
fedora-devel-list@redhat.com
https://www.redhat.com/mailman/listinfo/fedora-devel-list
 
Old 06-10-2008, 05:32 PM
Alan Cox
 
Default LVM negates benefits of jounaling filesystems?

On Tue, Jun 10, 2008 at 11:21:19AM -0500, Eric Sandeen wrote:
> quick test; fast hardware raid on fibrechannel, 4 cpu opteron, 8G
> memory, 4-threaded dbench for 600s, 120s warmup, 500G ext3 fs, 2.6.26-rc2.
>
> ext3 on /dev/sdc:
> Throughput 679.495 MB/sec 4 procs
>
> ext3 on /dev/mapper/myvg-lvol0 on /dev/sdc:
> Throughput 665.599 MB/sec 4 procs

Nice numbers, although I was testing pair of ATA disks

--
fedora-devel-list mailing list
fedora-devel-list@redhat.com
https://www.redhat.com/mailman/listinfo/fedora-devel-list
 
Old 06-10-2008, 08:07 PM
"max bianco"
 
Default LVM negates benefits of jounaling filesystems?

On Tue, Jun 10, 2008 at 9:46 AM, Chuck Anderson <cra@wpi.edu> wrote:
> On Tue, Jun 10, 2008 at 08:36:31AM -0500, Eric Sandeen wrote:
>> journaling filesystem you really shouldn't have any filesystem metadata
>> integrity problems on power loss; that is, if you have barriers on
>> (which ext3 doesn't by default) and if your storage can pass barriers
>> (which lvm doesn't), or if you have drive write cache disabled (which
>> hurts performance pretty badly).
>
> I wasn't aware that LVM destroyed the kind of guarantees about
> filesystem metadata being written out to disk that jounaling
> filesystems rely on? If so, should we perhaps rethink the decision to
> use LVM by default on Fedora installs?
>

What was the reason for using LVM in the first place. My most recent
install I was really tempted to not go with the defaults but because I
really don't know much about filesystems, I figured the best thing in
that case was stick to the defaults. Now I am reconsidering
again...could someone explain the comparative advantages/disadvantages
? Before i do something stupid .


--
If opinions were really like assholes we'd each have just one

--
fedora-devel-list mailing list
fedora-devel-list@redhat.com
https://www.redhat.com/mailman/listinfo/fedora-devel-list
 

Thread Tools




All times are GMT. The time now is 06:08 AM.

VBulletin, Copyright ©2000 - 2014, Jelsoft Enterprises Ltd.
Content Relevant URLs by vBSEO ©2007, Crawlability, Inc.
Copyright 2007 - 2008, www.linux-archive.org