2.6.19-rc5-mm1 progression

2.6.19-rc5-mm1 progression

[Larry Finger]

Since at least the 2.6.18 development kernels, my interactive sessions have sometimes been very 
sluggish whenever a cpu-intensive process is running. This has been so bad that the mouse is 
unresponsive. The system behaved as if it were swapping; however, swapfile usage is always 0. The 
system has 820 MB RAM. I have not reported this problem as I really couldn't quantify it due to its 
intermittent nature.

Recently, I needed to run 2.6.19-rc5-mm1 to check out a problem reported by a bcm43xx user. I found 
that even with a kernel build, a git pull, and a separate build running, the system was responsive 
for interactive tasks. I have not yet identified which of the -mm1 patches "fixes" the problem, but 
plan to do so. Is there the equivalent of 'git bisect' for the -mmX kernels?

Larry

Re: 2.6.19-rc5-mm1 progression

[Benoit Boissinot]

On 11/24/06, Larry Finger <Larry.Finger@lwfinger.net> wrote:
> Is there the equivalent of 'git bisect' for the -mmX kernels?
>
http://www.zip.com.au/~akpm/linux/patches/stuff/bisecting-mm-trees.txt

regards,

Benoit

Re: 2.6.19-rc5-mm1 progression

[Andrew Morton]

On Fri, 24 Nov 2006 17:36:27 +0100
"Benoit Boissinot" <bboissin@gmail.com> wrote:

> On 11/24/06, Larry Finger <Larry.Finger@lwfinger.net> wrote:
> > Is there the equivalent of 'git bisect' for the -mmX kernels?
> >
> http://www.zip.com.au/~akpm/linux/patches/stuff/bisecting-mm-trees.txt
> 

Please take the time to do that.  Yours is an interesting report - I'm not
aware of anything in there which was expected to cause a change of this
mature.

Re: 2.6.19-rc5-mm1 progression

[Larry Finger]

Andrew Morton wrote:
> On Fri, 24 Nov 2006 17:36:27 +0100
> "Benoit Boissinot" <bboissin@gmail.com> wrote:
> 
>> On 11/24/06, Larry Finger <Larry.Finger@lwfinger.net> wrote:
>>> Is there the equivalent of 'git bisect' for the -mmX kernels?
>>>
>> http://www.zip.com.au/~akpm/linux/patches/stuff/bisecting-mm-trees.txt
>>
> 
> Please take the time to do that.  Yours is an interesting report - I'm not
> aware of anything in there which was expected to cause a change of this
> mature.
> 

I am in the process of isolating the patch. I was already a user of quilt so that part of the 
learning process was no problem.

Larry

Re: 2.6.19-rc5-mm1 progression

[Larry Finger]

Andrew Morton wrote:
> On Fri, 24 Nov 2006 17:36:27 +0100
> "Benoit Boissinot" <bboissin@gmail.com> wrote:
> 
>> On 11/24/06, Larry Finger <Larry.Finger@lwfinger.net> wrote:
>>> Is there the equivalent of 'git bisect' for the -mmX kernels?
>>>
>> http://www.zip.com.au/~akpm/linux/patches/stuff/bisecting-mm-trees.txt
>>
> 
> Please take the time to do that.  Yours is an interesting report - I'm not
> aware of anything in there which was expected to cause a change of this
> mature.
> 

There are at least two patches in 2.6.19-rc5-mm2 that make my system much more responsive for 
interactive jobs. The one that has the majority of the effect is:

radix-tree-rcu-lockless-readside.patch

I have not been able to isolate the second patch, which has the lesser effect. All I can say is that 
it occurred before the above patch in patches/series. This patch was tested against 2.6.19 and fixed 
most of the problem on that version.

Larry Finger

Re: 2.6.19-rc5-mm1 progression

[Peter Zijlstra]

On Mon, 2006-12-04 at 21:32 -0600, Larry Finger wrote:
> Andrew Morton wrote:
> > On Fri, 24 Nov 2006 17:36:27 +0100
> > "Benoit Boissinot" <bboissin@gmail.com> wrote:
> > 
> >> On 11/24/06, Larry Finger <Larry.Finger@lwfinger.net> wrote:
> >>> Is there the equivalent of 'git bisect' for the -mmX kernels?
> >>>
> >> http://www.zip.com.au/~akpm/linux/patches/stuff/bisecting-mm-trees.txt
> >>
> > 
> > Please take the time to do that.  Yours is an interesting report - I'm not
> > aware of anything in there which was expected to cause a change of this
> > mature.
> > 
> 
> There are at least two patches in 2.6.19-rc5-mm2 that make my system much more responsive for 
> interactive jobs. The one that has the majority of the effect is:
> 
> radix-tree-rcu-lockless-readside.patch
> 
> I have not been able to isolate the second patch, which has the lesser effect. All I can say is that 
> it occurred before the above patch in patches/series. This patch was tested against 2.6.19 and fixed 
> most of the problem on that version.

Curious...

This patch introduces the direct pointer optimisation for single element
radix trees and makes the radix tree safe to read in a lock-less manner
which is not used -yet-. The only difference that that should have is
that the elements are freed using rcu callback instead of directly.

/me puzzled how this has a large effect on interactivity.

Nick?

Re: 2.6.19-rc5-mm1 progression

[Larry Finger]

Peter Zijlstra wrote:
> On Mon, 2006-12-04 at 21:32 -0600, Larry Finger wrote:
>> Andrew Morton wrote:
>>> On Fri, 24 Nov 2006 17:36:27 +0100
>>> "Benoit Boissinot" <bboissin@gmail.com> wrote:
>>>
>>>> On 11/24/06, Larry Finger <Larry.Finger@lwfinger.net> wrote:
>>>>> Is there the equivalent of 'git bisect' for the -mmX kernels?
>>>>>
>>>> http://www.zip.com.au/~akpm/linux/patches/stuff/bisecting-mm-trees.txt
>>>>
>>> Please take the time to do that.  Yours is an interesting report - I'm not
>>> aware of anything in there which was expected to cause a change of this
>>> mature.
>>>
>> There are at least two patches in 2.6.19-rc5-mm2 that make my system much more responsive for 
>> interactive jobs. The one that has the majority of the effect is:
>>
>> radix-tree-rcu-lockless-readside.patch
>>
>> I have not been able to isolate the second patch, which has the lesser effect. All I can say is that 
>> it occurred before the above patch in patches/series. This patch was tested against 2.6.19 and fixed 
>> most of the problem on that version.
> 
> Curious...
> 
> This patch introduces the direct pointer optimisation for single element
> radix trees and makes the radix tree safe to read in a lock-less manner
> which is not used -yet-. The only difference that that should have is
> that the elements are freed using rcu callback instead of directly.
> 
> /me puzzled how this has a large effect on interactivity.

I also wondered why this patch should have the effect as it didn't seem in an area that should be 
important. I did test it on both 2.6.19-rc5-mm2 and vanilla 2.6.19. In both cases, it made 
interactive response improve.

One single task that seems to have the biggest effect on my unpatched systems is the link step for a 
new kernel. I am assuming that this process calls for lots of virtual memory. One other 
characteristic of my system when it gets unresponsive is that kswapd is running regularly; however, 
I have never observed any swap file usage.

Larry

Re: 2.6.19-rc5-mm1 progression

[Nick Piggin]

Peter Zijlstra wrote:
> On Mon, 2006-12-04 at 21:32 -0600, Larry Finger wrote:

>>There are at least two patches in 2.6.19-rc5-mm2 that make my system much more responsive for 
>>interactive jobs. The one that has the majority of the effect is:
>>
>>radix-tree-rcu-lockless-readside.patch
>>
>>I have not been able to isolate the second patch, which has the lesser effect. All I can say is that 
>>it occurred before the above patch in patches/series. This patch was tested against 2.6.19 and fixed 
>>most of the problem on that version.
> 
> 
> Curious...
> 
> This patch introduces the direct pointer optimisation for single element
> radix trees and makes the radix tree safe to read in a lock-less manner
> which is not used -yet-. The only difference that that should have is
> that the elements are freed using rcu callback instead of directly.
> 
> /me puzzled how this has a large effect on interactivity.
> 
> Nick?

I have already got the direct data optimisation upstream. It might be
possible that it is some interaction with the extra rcu callbacks going
off... I don't know :\

-- 
SUSE Labs, Novell Inc.
Send instant messages to your online friends http://au.messenger.yahoo.com