Home » Mailing lists » Devel » Re: [ckrm-tech] [RFC][PATCH 5/7] UBC: kernel memory accounting (core)
| Re: [ckrm-tech] [RFC][PATCH 5/7] UBC: kernel memory accounting (core) [message #5229] |
Wed, 16 August 2006 19:59  |
Dave Hansen
Messages: 240
Registered: October 2005
|
Senior Member |
|
|
On Wed, 2006-08-16 at 12:15 -0700, Rohit Seth wrote:
> My preference would be to have container (I keep on saying container,
> but resource beancounter) pointer embeded in task, mm(not sure),
> address_space and anon_vma structures.
Hmm. If we can embed it in the mm, then we can get there from any given
anon_vma (or any pte for that matter). Here's a little prototype for
doing just that:
http://www.sr71.net/patches/2.6.18/2.6.18-rc4-mm1-lxc1/broke n-out/modify-lru-walk.patch
See file/anon_page_has_naughty_cpuset(). Anybody see any basic problems
with doing it that way?
One trick with putting it in an mm is that we don't have a direct
relationship between processes and mm's. We could also potentially have
two different threads of a process in two different accounting contexts.
But, that might be as simple to fix as disallowing things that share mms
from being in different accounting contexts, unless you unshare the mm.
-- Dave
|
|
|
|
|
|
| Re: [ckrm-tech] [RFC][PATCH 5/7] UBC: kernel memory accounting (core) [message #5319 is a reply to message #5242] |
Thu, 17 August 2006 14:26  |
Dave Hansen
Messages: 240
Registered: October 2005
|
Senior Member |
|
|
On Thu, 2006-08-17 at 01:24 +0100, Alan Cox wrote:
> Ar Mer, 2006-08-16 am 12:59 -0700, ysgrifennodd Dave Hansen:
> > relationship between processes and mm's. We could also potentially have
> > two different threads of a process in two different accounting contexts.
> > But, that might be as simple to fix as disallowing things that share mms
> > from being in different accounting contexts, unless you unshare the mm.
>
> At the point I have twenty containers containing 20 copies of glibc to
> meet your suggestion it would be *far* cheaper to put it in the page
> struct.
My main thought is that _everybody_ is going to have to live with the
entry in the 'struct page'. Distros ship one kernel for everybody, and
the cost will be paid by those not even using any kind of resource
control or containers.
That said, it sure is simpler to implement, so I'm all for it!
-- Dave
|
|
|
|
|
|
| Re: [ckrm-tech] [RFC][PATCH 5/7] UBC: kernel memory accounting (core) [message #5324 is a reply to message #5229] |
Thu, 17 August 2006 16:31 |
Rohit Seth
Messages: 101
Registered: August 2006
|
Senior Member |
|
|
On Wed, 2006-08-16 at 12:59 -0700, Dave Hansen wrote:
> On Wed, 2006-08-16 at 12:15 -0700, Rohit Seth wrote:
> > My preference would be to have container (I keep on saying container,
> > but resource beancounter) pointer embeded in task, mm(not sure),
> > address_space and anon_vma structures.
>
> Hmm. If we can embed it in the mm, then we can get there from any given
> anon_vma (or any pte for that matter). Here's a little prototype for
> doing just that:
>
> http://www.sr71.net/patches/2.6.18/2.6.18-rc4-mm1-lxc1/broke n-out/modify-lru-walk.patch
>
> See file/anon_page_has_naughty_cpuset(). Anybody see any basic problems
> with doing it that way?
>
> One trick with putting it in an mm is that we don't have a direct
> relationship between processes and mm's. We could also potentially have
> two different threads of a process in two different accounting contexts.
> But, that might be as simple to fix as disallowing things that share mms
> from being in different accounting contexts, unless you unshare the mm.
But anon_vmas could be shared across different processes (with different
mms).
-rohit
|
|
|
|
|
|
| Re: [ckrm-tech] [RFC][PATCH 5/7] UBC: kernel memory accounting (core) [message #5333 is a reply to message #5319] |
Thu, 17 August 2006 17:16 |
Rohit Seth
Messages: 101
Registered: August 2006
|
Senior Member |
|
|
On Thu, 2006-08-17 at 07:26 -0700, Dave Hansen wrote:
> On Thu, 2006-08-17 at 01:24 +0100, Alan Cox wrote:
> > Ar Mer, 2006-08-16 am 12:59 -0700, ysgrifennodd Dave Hansen:
> > > relationship between processes and mm's. We could also potentially have
> > > two different threads of a process in two different accounting contexts.
> > > But, that might be as simple to fix as disallowing things that share mms
> > > from being in different accounting contexts, unless you unshare the mm.
> >
> > At the point I have twenty containers containing 20 copies of glibc to
> > meet your suggestion it would be *far* cheaper to put it in the page
> > struct.
>
> My main thought is that _everybody_ is going to have to live with the
> entry in the 'struct page'. Distros ship one kernel for everybody, and
> the cost will be paid by those not even using any kind of resource
> control or containers.
>
> That said, it sure is simpler to implement, so I'm all for it!
hmm, not sure why it is simpler.
-rohit
|
|
|
|
|
|
|
|
|
|
| Re: [ckrm-tech] [RFC][PATCH 5/7] UBC: kernel memory accounting (core) [message #5344 is a reply to message #5333] |
Thu, 17 August 2006 17:23 |
Dave Hansen
Messages: 240
Registered: October 2005
|
Senior Member |
|
|
On Thu, 2006-08-17 at 10:16 -0700, Rohit Seth wrote:
> > That said, it sure is simpler to implement, so I'm all for it!
>
> hmm, not sure why it is simpler.
When you ask the question, "which container owns this page?", you don't
have to look far, nor is it ambiguous in any way. It is very strict,
and very straightforward.
-- Dave
|
|
|
|
|
|
| Re: [ckrm-tech] [RFC][PATCH 5/7] UBC: kernel memory accounting (core) [message #5366 is a reply to message #5333] |
Fri, 18 August 2006 08:49 |
dev
Messages: 1693
Registered: September 2005
Location: Moscow
|
Senior Member |
|
|
Rohit Seth wrote:
> On Thu, 2006-08-17 at 07:26 -0700, Dave Hansen wrote:
>
>>On Thu, 2006-08-17 at 01:24 +0100, Alan Cox wrote:
>>
>>>Ar Mer, 2006-08-16 am 12:59 -0700, ysgrifennodd Dave Hansen:
>>>
>>>>relationship between processes and mm's. We could also potentially have
>>>>two different threads of a process in two different accounting contexts.
>>>>But, that might be as simple to fix as disallowing things that share mms
>>>>from being in different accounting contexts, unless you unshare the mm.
>>>
>>>At the point I have twenty containers containing 20 copies of glibc to
>>>meet your suggestion it would be *far* cheaper to put it in the page
>>>struct.
>>
>>My main thought is that _everybody_ is going to have to live with the
>>entry in the 'struct page'. Distros ship one kernel for everybody, and
>>the cost will be paid by those not even using any kind of resource
>>control or containers.
>>
>>That said, it sure is simpler to implement, so I'm all for it!
>
>
>
> hmm, not sure why it is simpler.
because introducing additonal lookups/hashes etc. is harder and
adds another source for possible mistakes.
we can always optimize it out if people insist.
Kirill
|
|
|
|
| Re: [ckrm-tech] [RFC][PATCH 5/7] UBC: kernel memory accounting (core) [message #5367 is a reply to message #5333] |
Fri, 18 August 2006 08:49 |
dev
Messages: 1693
Registered: September 2005
Location: Moscow
|
Senior Member |
|
|
Rohit Seth wrote:
> On Thu, 2006-08-17 at 07:26 -0700, Dave Hansen wrote:
>
>>On Thu, 2006-08-17 at 01:24 +0100, Alan Cox wrote:
>>
>>>Ar Mer, 2006-08-16 am 12:59 -0700, ysgrifennodd Dave Hansen:
>>>
>>>>relationship between processes and mm's. We could also potentially have
>>>>two different threads of a process in two different accounting contexts.
>>>>But, that might be as simple to fix as disallowing things that share mms
>>>>from being in different accounting contexts, unless you unshare the mm.
>>>
>>>At the point I have twenty containers containing 20 copies of glibc to
>>>meet your suggestion it would be *far* cheaper to put it in the page
>>>struct.
>>
>>My main thought is that _everybody_ is going to have to live with the
>>entry in the 'struct page'. Distros ship one kernel for everybody, and
>>the cost will be paid by those not even using any kind of resource
>>control or containers.
>>
>>That said, it sure is simpler to implement, so I'm all for it!
>
>
>
> hmm, not sure why it is simpler.
because introducing additonal lookups/hashes etc. is harder and
adds another source for possible mistakes.
we can always optimize it out if people insist (by cost of slower accounting).
Kirill
|
|
|
|
|
|
| Re: [ckrm-tech] [RFC][PATCH 5/7] UBC: kernel memory accounting (core) [message #5412 is a reply to message #5362] |
Fri, 18 August 2006 17:06 |
Rohit Seth
Messages: 101
Registered: August 2006
|
Senior Member |
|
|
On Fri, 2006-08-18 at 12:29 +0400, Kirill Korotaev wrote:
> Rik van Riel wrote:
> > Dave Hansen wrote:
> >
> >> My main thought is that _everybody_ is going to have to live with the
> >> entry in the 'struct page'. Distros ship one kernel for everybody, and
> >> the cost will be paid by those not even using any kind of resource
> >> control or containers.
> >
> >
> > Every userspace or page cache page will be in an object
> > though. Could we do the pointer on a per object (mapping,
> > anon vma, ...) basis?
> in this case no memory fractions accounting is possible :/
> please, note, this field added by this patchset is in union
> and used by user pages accounting as well.
>
> > Kernel pages are not using all of their struct page entries,
> > so we could overload a field.
> yeah, we can. probably mapping.
> but as I said we use the same pointer for user pages accounting as well.
>
> > It all depends on how much we really care about not growing
> > struct page :)
> so what is your opinion?
> Kernel compiled w/o UBC do not introduce additional pointer.
As Andi pointed out earlier that slab and network codes are going to use
the mapping field (and you pointed out that some of this is allocated
out of context), so seems like for kernel accounting we will need
another field in page structure.
-rohit
|
|
|
|
| Re: [ckrm-tech] [RFC][PATCH 5/7] UBC: kernel memory accounting (core) [message #5422 is a reply to message #5400] |
Fri, 18 August 2006 19:32 |
Dave Hansen
Messages: 240
Registered: October 2005
|
Senior Member |
|
|
On Fri, 2006-08-18 at 15:59 +0100, Alan Cox wrote:
> Ar Gwe, 2006-08-18 am 12:52 +0400, ysgrifennodd Kirill Korotaev:
> > > hmm, not sure why it is simpler.
> > because introducing additonal lookups/hashes etc. is harder and
> > adds another source for possible mistakes.
> > we can always optimize it out if people insist (by cost of slower accounting).
>
> It ought to be cheap. Given each set of page structs is an array its a
> simple subtract and divide (or with care and people try to pack them
> nicely for cache lines - shift) to get to the parallel accounting array.
I wish page structs were just a simple array. ;)
It will just be a bit more code, but we'll need this for the two other
memory models: sparsemem and discontigmem. For discontig, we'll just
need pointers in the pg_data_ts and, for sparsemem, we'll likely need
another pointer in the 'struct mem_section'.
This will effectively double the memory we need for sparsemem (because
we only use one pointer per SECTION_SIZE bytes of memory) but, that
should be just fine.
Is there ever any need to go from the accounting structure *back* to the
page? I guess that might be the hard part with keeping parallel arrays,
if we even need it.
The reverse lookups might introduce a bit more pain with sparsemem and
discontig because, right now, we use bits in page->flags to help us go
find the containing node or the correct mem_section for the page.
-- Dave
|
|
|
|
|
|
|
|
Goto Forum:
Current Time: Fri Oct 18 12:42:52 GMT 2024
Total time taken to generate the page: 0.11323 seconds
|
OpenVZ Forum
Members
Search
Help
Register
Login
Home
