Re: [RFC 2.6.11-rc2-mm2 7/7] mm: manual page migration -- sys_page_migrate

["Martin J. Bligh" <mbligh@aracnet.com>]

>>> I think that part of the motivation here (e. g. the batch scheduler on
>>> a  large NUMA machine) is to push pages off of the old nodes so that
>>> a new job running on the old nodes can allocate memory quickly and
>>> efficiently (i. e. without having to swap out the old job's pages).
>> 
>> 
>> If our VM code wasn't crap, we'd do that automatically. It seems somewhat
>> excessive to do that from a manual interface?
>> 
> 
> SGI had code in IRIX to do that kind of thing (automatically move a page to
> the node where most of the references were coming from).  Never worked very well, I have been told.   So our bias is away from such "automatic" page
> migration schemes and toward "manual" methods driven either by a user
> command or a user-level program such as a batch scheduler.

I realize it's a non-trivial problem. But if we don't tackle it, at least
at some basic level, then machines won't run will without manual tweaking
up the wazoo, which isn't where most of us want to end up.

>>> True enough, we may move pages that are not currently being used.
>>> But. on our large NUMA systems, we want the nodes where a new job
>>> starts to be relatively clean so that local page allocations are
>>> indeed satisfied by local pages and that these requests do not
>>> spill off node.
>> 
>> 
>> Yes. The objective was to kick the LRU page off this node onto some other
>> node, or to disk ... at the moment, if one node is more heavily used, we
>> will always allocate off node for all new pages. that's crap.
> 
> Tell me about it.  I've spent most of the past couple of years wrassling
> with VM system to get it to behave in this situation.  :-)

If the workload is fairly balanced, we'd be better off on some machines 
just not using the fallback, and getting local allocation. Quite how we
recognise when that's the right thing to do is a bit of a mystery.

I suppose I'd better take one of our more modern boxes, and try benchmarking
it. I have a 4x Opteron somewhere.

M.

--
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:"aart@kvack.org"> aart@kvack.org </a>