[prev in list] [next in list] [prev in thread] [next in thread] 

List:       linux-ext4
Subject:    Re: Ext4 performance regression: Post 2.6.30
From:       Greg Freemyer <greg.freemyer () gmail ! com>
Date:       2010-03-31 22:55:17
Message-ID: s2v87f94c371003311555i2cecadd2y4906eb5b98f40981 () mail ! gmail ! com
[Download RAW message or body]

On Wed, Mar 31, 2010 at 6:14 PM, Keith Mannthey <kmannth@us.ibm.com> wrote:
> On Wed, 2010-03-31 at 18:06 -0400, Greg Freemyer wrote:
> > On Wed, Mar 31, 2010 at 6:02 PM, Keith Mannthey <kmannth@us.ibm.com> wrote:
> > > On Tue, 2010-03-30 at 23:06 -0500, Eric Sandeen wrote:
> > > > Keith Mannthey wrote:
> > > > > On Mon, 2010-03-29 at 11:10 -0400, Greg Freemyer wrote:
> > > > > > On Mon, Mar 29, 2010 at 2:25 AM, Keith Mannthey <kmannth@us.ibm.com> \
> > > > > > wrote:
> > > > > > > 
> > > > > > > After 2.6.30 I am seeing large performance regressions on a raid setup.
> > > > > > > I am working to publish a larger amount of data but I wanted to get \
> > > > > > > some quick data out about what I am seeing.
> > > > > > > 
> > > > > > Is mdraid involved?
> > > > > > 
> > > > > > They added barrier support for some configs after 2.6.30 I believe.
> > > > > > It can cause a drastic perf change, but it increases reliability and
> > > > > > is "correct".
> > > > > 
> > > > > lvm and device mapper are is involved.  The git bisect just took me to:
> > > > > 
> > > > > 374bf7e7f6cc38b0483351a2029a97910eadde1b is first bad commit
> > > > > commit 374bf7e7f6cc38b0483351a2029a97910eadde1b
> > > > > Author: Mikulas Patocka <mpatocka@redhat.com>
> > > > > Date:   Mon Jun 22 10:12:22 2009 +0100
> > > > > 
> > > > > dm: stripe support flush
> > > > > 
> > > > > Flush support for the stripe target.
> > > > > 
> > > > > This sets ti->num_flush_requests to the number of stripes and
> > > > > remaps individual flush requests to the appropriate stripe devices.
> > > > > 
> > > > > Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
> > > > > Signed-off-by: Alasdair G Kergon <agk@redhat.com>
> > > > > 
> > > > > > 040000 040000 542f4b9b442d1371c6534f333b7e00714ef98609 \
> > > > > > d490479b660139fc1b6b0ecd17bb58c9e00e597e M  drivers
> > > > > 
> > > > > 
> > > > > This may be correct behavior but the performance penalty in this test
> > > > > case is pretty high.
> > > > > 
> > > > > I am going to move back to current kernels and starting looking into
> > > > > ext4/dm flushing.
> > > > 
> > > > It would probably be interesting to do a mount -o nobarrier to see if
> > > > that makes the regression go away.
> > > 
> > > -o nobarrier takes the regression away with 2.6.34-rc3:
> > > 
> > > Default mount: ~27500
> > > 
> > > -o nobarrier: ~12500
> > > 
> > > Barriers on this setup cost ALOT during writes.
> > > 
> > > Interestingly as well the "mailserver" workload regression is also
> > > removed by mounting with "-o nobarrier".
> > > 
> > > I am going to see what impact is seen on a single disk setup.
> > > 
> > > Thanks,
> > > Keith Mannthey
> > > LTC FS-Dev
> > 
> > I'm curious if your using an internal or external journal?
> 
> I am unsure.  How do I tell?  I am using defaults except with the -o
> nobarrier.   I know jdb2 is being used.
> 
> Thanks,
> Keith

The default is internal.  External requires a separate partition be
provided to hold the journal.

Since journals are typically very small relative to the overall
filesystem, a small raid 1 partition would be my production
recommendation to hold an external journal.

But for performance testing purposes, if you have a drive that is not
participating in your current raid setup, you can simply create a
small partition on it and use it to hold the external journal.  I
believe you can convert your existing file system to an external
journal easily and without having to recreate your file system.

Greg
--
To unsubscribe from this list: send the line "unsubscribe linux-ext4" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html


[prev in list] [next in list] [prev in thread] [next in thread]