> On Mon, 2008-05-26 at 15:33 -0400, Rik van Riel wrote:
>> On Mon, 26 May 2008 23:54:55 +0530
>> Balbir Singh <balbir@linux.vnet.ibm.com> wrote:
>>
>>> Rik van Riel wrote:
>>>> On large memory systems, the VM can spend way too much time scanning
>>>> through pages that it cannot (or should not) evict from memory. Not
>>>> only does it use up CPU time, but it also provokes lock contention
>>>> and can leave large systems under memory presure in a catatonic state.
>>> Hi, Rik,
>>>
>>> This patchset looks good (I did a brief scan). I'll go ahead and play with it?
>>> What is a good memory size to test the patches on (to see improvements).
>> The larger, the better. One known problem with the current upstream
>> VM is large numbers of anonymous pages, or a mix of mlocked and anon
>> pages.
>>
>> Once the system needs to swap something out, every single anon page
>> will have the referenced bit set and the system needs to do lots of
>> scanning before it can evict the first page. This scanning causes
>> multiple CPUs to pile up and things slow down exponentially and/or
>> catastrophically :)
>>
>> Unfortunately the largest system I have access to on a regular basis
>> has "only" 16GB of RAM :(
>>
>> I am also making 2.6.25 based kernel RPMs available with the split LRU
>> patch set, at
http://people.redhat.com/riel/splitvm/
>>
>> The most recently posted patches are newer, though...
>>
>
> I tested Rik's previous patch set with my noreclaim/mlock patches over
> the long weekend on 32GB systems--one ia64 [16cpu x 4 nodes] and one
> x86_64 [8 core x 4 nodes] on 26-rc2-mm1. A fairly heavy stress load ran
> for 92-93 hours on each system w/o error. Stats tracked throughout, no
> leaked pages, ...
>
> Since Balbir is starting to look at this, I need to ask about
> interaction with the memory controller. It is currently unaware of the
> noreclaim list. I'm not sure what will happen if/when the memory
> controller tries to reclaim a page that system has moved to the
> noreclaim list. Something we'll need to address. It's on my list, but
> I won't get to it for a couple of weeks.
