Re: Major performance regressions in 3.7rc1/2

Re: Major performance regressions in 3.7rc1/2

[Andrew Morton]

On Mon, 22 Oct 2012 21:45:02 +0200
Julian Wollrath <julian.wollrath@stud.uni-goettingen.de> wrote:

> Hello,
> 
> seems like I found the other bad commit. Everything, which means
> v3.7-rc*, works fine again with commit e6c509f85 (mm: use
> clear_page_mlock() in page_remove_rmap()) and commit 957f822a0 (mm,
> numa: reclaim from all nodes within reclaim distance) revoked.

Thanks.  Let's add some cc's.  Can you please describe your workload
and some estimate of the slowdown?

Patrik has also seen this and his description is

: I'm using an Asus Zenbook UX31E and have been installing all RCs in
: hope of improving the Wireless and Touchpad functionality.
: However, when trying 3.7 (rc1 and now rc2) I have major performance issues.
: 
: Easiest way to reproduce is to launch and play a game like Nexuiz,
: where the computer will lag, stutter and freeze until the machine is
: unresponsive within a couple of minutes.
: But an easy workload like browsing will also cause lags when switching
: tabs or redrawing a web page after a tab switch.
: Basically 3.7 is unusable for this machine.

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

Re: Major performance regressions in 3.7rc1/2

[Hugh Dickins]

On Mon, 22 Oct 2012, Andrew Morton wrote:
> On Mon, 22 Oct 2012 21:45:02 +0200
> Julian Wollrath <julian.wollrath@stud.uni-goettingen.de> wrote:
> 
> > Hello,
> > 
> > seems like I found the other bad commit. Everything, which means
> > v3.7-rc*, works fine again with commit e6c509f85 (mm: use
> > clear_page_mlock() in page_remove_rmap()) and commit 957f822a0 (mm,
> > numa: reclaim from all nodes within reclaim distance) revoked.
> 
> Thanks.  Let's add some cc's.  Can you please describe your workload
> and some estimate of the slowdown?

I'm currently assuming that my clear_page_mlock() commit is innocent
of this: it went in just two before David's numa reclaim commit, and
I don't see how mine could have any such marked effect: I'm thinking
it was just a bisection hiccup that implicated it.

Julian (or Patrik), what do you find if you revert the numa reclaim
commit but leave in the clear_page_mlock commit?  Or vice versa.

Whereas David's numa reclaim distance commit does look suspicious to
me: as I read it (perhaps wrongly), zone_reclaim_mode would previously
have stayed 0 if all nodes were close or equidistant (e.g. CONFIG_NUMA=y
on a non-NUMA machine), but now it would be 1 in this case, invoking
zone_reclaim() more frequently than before.

Hugh

> 
> Patrik has also seen this and his description is
> 
> : I'm using an Asus Zenbook UX31E and have been installing all RCs in
> : hope of improving the Wireless and Touchpad functionality.
> : However, when trying 3.7 (rc1 and now rc2) I have major performance issues.
> : 
> : Easiest way to reproduce is to launch and play a game like Nexuiz,
> : where the computer will lag, stutter and freeze until the machine is
> : unresponsive within a couple of minutes.
> : But an easy workload like browsing will also cause lags when switching
> : tabs or redrawing a web page after a tab switch.
> : Basically 3.7 is unusable for this machine.

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

Re: Major performance regressions in 3.7rc1/2

[Julian Wollrath]

> > Thanks.  Let's add some cc's.  Can you please describe your workload
> > and some estimate of the slowdown?
I am using fluxbox with Iceweasel, Claws-Mail and urxvt on different
workspaces on a Thinkpad X121e with an AMD E-450 APU. Loading some big
pages in Iceweasel leades to a very sluggish rendering of the urxvt
window when changing workspaces, the cursor movement falters. The
falter in the cursor movement is from random length but I would
estimate, that it is mostly under one second. But sometimes the time
between the each falter is very short which results in a more or less
unusable system.

> I'm currently assuming that my clear_page_mlock() commit is innocent
> of this: it went in just two before David's numa reclaim commit, and
> I don't see how mine could have any such marked effect: I'm thinking
> it was just a bisection hiccup that implicated it.
Just tested v3.7-rc2 with your clear_page_mlock() and without the numa
reclaim commit and everything worked fine. So you are right, most
probable it was a bisection hiccup, the reclaim commit is the real bad
commit. Nevertheless I am wondering why everything worked fine until
39b5f29a (mm: remove vma arg from page_evictable) and then started to
behave badly with your clear_page_mlock() commit but 3.7-rc2 works fine
with only the numa reclaim commit revoked.


With best regards,
Julian Wollrath

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

Re: Major performance regressions in 3.7rc1/2

[Anca Emanuel]

I have the same problem.
Reverting
https://github.com/torvalds/linux/commit/957f822a0ab95e88b146638bad6209bbc315bedd
solves the problem for me.

Here is dmesg: http://pastebin.com/r78Rcrf5

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

Re: Major performance regressions in 3.7rc1/2

[David Rientjes]

On Tue, 23 Oct 2012, Anca Emanuel wrote:

> I have the same problem.
> Reverting
> https://github.com/torvalds/linux/commit/957f822a0ab95e88b146638bad6209bbc315bedd
> solves the problem for me.
> 

If you don't revert anything and do

	echo 0 > /proc/sys/vm/zone_reclaim_mode

after boot, does this also fix the issue?

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

Re: Major performance regressions in 3.7rc1/2

[Anca Emanuel]

On Wed, Oct 24, 2012 at 1:42 AM, David Rientjes <rientjes@google.com> wrote:
> On Tue, 23 Oct 2012, Anca Emanuel wrote:
>
>> I have the same problem.
>> Reverting
>> https://github.com/torvalds/linux/commit/957f822a0ab95e88b146638bad6209bbc315bedd
>> solves the problem for me.
>>
>
> If you don't revert anything and do
>
>         echo 0 > /proc/sys/vm/zone_reclaim_mode
>
> after boot, does this also fix the issue?

Yes.
http://imgur.com/JJwiJ

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

[patch for-3.7] mm, numa: avoid setting zone_reclaim_mode unless a node is sufficiently distant

[David Rientjes]

Commit 957f822a0ab9 ("mm, numa: reclaim from all nodes within reclaim
distance") caused zone_reclaim_mode to be set for all systems where two
nodes are within RECLAIM_DISTANCE of each other.  This is the opposite of
what we actually want: zone_reclaim_mode should be set if two nodes are
sufficiently distant.

Reported-by: Julian Wollrath <jwollrath@web.de>
Signed-off-by: David Rientjes <rientjes@google.com>
---
 mm/page_alloc.c |    4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/mm/page_alloc.c b/mm/page_alloc.c
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c
@@ -1809,10 +1809,10 @@ static void __paginginit init_zone_allows_reclaim(int nid)
 	int i;
 
 	for_each_online_node(i)
-		if (node_distance(nid, i) <= RECLAIM_DISTANCE) {
+		if (node_distance(nid, i) <= RECLAIM_DISTANCE)
 			node_set(i, NODE_DATA(nid)->reclaim_nodes);
+		else
 			zone_reclaim_mode = 1;
-		}
 }
 
 #else	/* CONFIG_NUMA */

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

Re: [patch for-3.7] mm, numa: avoid setting zone_reclaim_mode unless a node is sufficiently distant

[Julian Wollrath]

This patch fixes the problem for me, thank you. Feal free to add a 

Tested-by: Julian Wollrath <jwollrath@web.de>


With best regards,
Julian Wollrath

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>