multi-second freezes with current GIT?

multi-second freezes with current GIT?

[David Miller]

I've been noticing this off and on for the past week or so.

The system seems to jam up for several seconds, anything that would
need to read from disk just sits there during this time.  I think it's
correlated with generating a lot of dirty write data.

My mouse moves around in X etc. so it really is only processes that
need to read from disk that get stuck.

Perhaps it's a side effect of those dirty ratio changes Linus made to
start off the 2.6.22 merge cycle?

Re: multi-second freezes with current GIT?

[Mike Galbraith]

On Tue, 2007-05-08 at 18:44 -0700, David Miller wrote:
> I've been noticing this off and on for the past week or so.
> 
> The system seems to jam up for several seconds, anything that would
> need to read from disk just sits there during this time.  I think it's
> correlated with generating a lot of dirty write data.
> 
> My mouse moves around in X etc. so it really is only processes that
> need to read from disk that get stuck.
> 
> Perhaps it's a side effect of those dirty ratio changes Linus made to
> start off the 2.6.22 merge cycle?

Uhoh, that seems highly likely.

I think those changes may have been triggered by my repeatable and truly
horrible system stalls when writing to an ext3 data=ordered nearly full
filesystem.  Anything that does fdatasync() and/or fsync() can cause
very bad experiences indeed.  KDE's little menu/program launcher
doohickey does fdatasync() for some odd reason, which has utterly killed
my entire GUI for up to and including 20 minutes at a whack.

Switching to data=writeback cured those horrors.  Lowering the dirty
ratio reduced the agony of data=ordered tremendously, but didn't make it
even remotely acceptable.

	-Mike

Re: multi-second freezes with current GIT?

[David Miller]

From: Mike Galbraith <efault@gmx.de>
Date: Wed, 09 May 2007 05:29:20 +0200

> On Tue, 2007-05-08 at 18:44 -0700, David Miller wrote:
> > I've been noticing this off and on for the past week or so.
> > 
> > The system seems to jam up for several seconds, anything that would
> > need to read from disk just sits there during this time.  I think it's
> > correlated with generating a lot of dirty write data.
> > 
> > My mouse moves around in X etc. so it really is only processes that
> > need to read from disk that get stuck.
> > 
> > Perhaps it's a side effect of those dirty ratio changes Linus made to
> > start off the 2.6.22 merge cycle?
> 
> Uhoh, that seems highly likely.
> 
> I think those changes may have been triggered by my repeatable and truly
> horrible system stalls when writing to an ext3 data=ordered nearly full
> filesystem.  Anything that does fdatasync() and/or fsync() can cause
> very bad experiences indeed.  KDE's little menu/program launcher
> doohickey does fdatasync() for some odd reason, which has utterly killed
> my entire GUI for up to and including 20 minutes at a whack.
> 
> Switching to data=writeback cured those horrors.  Lowering the dirty
> ratio reduced the agony of data=ordered tremendously, but didn't make it
> even remotely acceptable.

I'm using data=ordered on my partitions too.

Re: multi-second freezes with current GIT?

[Andrew Morton]

On Tue, 08 May 2007 23:23:26 -0700 (PDT) David Miller <davem@davemloft.net> wrote:

> From: Mike Galbraith <efault@gmx.de>
> Date: Wed, 09 May 2007 05:29:20 +0200
> 
> > On Tue, 2007-05-08 at 18:44 -0700, David Miller wrote:
> > > I've been noticing this off and on for the past week or so.
> > > 
> > > The system seems to jam up for several seconds, anything that would
> > > need to read from disk just sits there during this time.  I think it's
> > > correlated with generating a lot of dirty write data.
> > > 
> > > My mouse moves around in X etc. so it really is only processes that
> > > need to read from disk that get stuck.
> > > 
> > > Perhaps it's a side effect of those dirty ratio changes Linus made to
> > > start off the 2.6.22 merge cycle?
> > 
> > Uhoh, that seems highly likely.
> > 
> > I think those changes may have been triggered by my repeatable and truly
> > horrible system stalls when writing to an ext3 data=ordered nearly full
> > filesystem.  Anything that does fdatasync() and/or fsync() can cause
> > very bad experiences indeed.  KDE's little menu/program launcher
> > doohickey does fdatasync() for some odd reason, which has utterly killed
> > my entire GUI for up to and including 20 minutes at a whack.
> > 
> > Switching to data=writeback cured those horrors.  Lowering the dirty
> > ratio reduced the agony of data=ordered tremendously, but didn't make it
> > even remotely acceptable.
> 
> I'm using data=ordered on my partitions too.


does

echo 40 > /proc/sys/vm/dirty_ratio
echo 10 > /proc/sys/vm/dirty_background_ratio

fix it?

Re: multi-second freezes with current GIT?

[David Miller]

From: Andrew Morton <akpm@linux-foundation.org>
Date: Wed, 9 May 2007 01:43:28 -0700

> does
> 
> echo 40 > /proc/sys/vm/dirty_ratio
> echo 10 > /proc/sys/vm/dirty_background_ratio
> 
> fix it?

Yes, I was going to play with that, I'll report if anything
decisive happens.