Re: [PATCH -mm] do_migrate_pages() calls migrate_to_node() even if task is already on a correct node

[KOSAKI Motohiro <mkosaki@redhat.com>]

(3/22/12 2:51 PM), Christoph Lameter wrote:
> On Thu, 22 Mar 2012, KOSAKI Motohiro wrote:
>
>> CC to Christoph.
>>
>>> While moving tasks between cpusets I noticed some strange behavior.
>>> Specifically if the nodes of the destination
>>> cpuset are a subset of the nodes of the source cpuset do_migrate_pages()
>>> will move pages that are already on a node
>>> in the destination cpuset. The reason for this is do_migrate_pages() does
>>> not check whether each node in the source
>>> nodemask is in the destination nodemask before calling migrate_to_node(). If
>>> we simply do this check and skip them
>>> when the source is in the destination moving we wont move nodes that dont
>>> need to be moved.
>>>
>>> Adding a little debug printk to migrate_to_node():
>>>
>>> Without this change migrating tasks from a cpuset containing nodes 0-7 to a
>>> cpuset containing nodes 3-4, we migrate
>>> from ALL the nodes even if they are in the both the source and destination
>>> nodesets:
>>>
>>> Migrating 7 to 4
>>> Migrating 6 to 3
>>> Migrating 5 to 4
>>> Migrating 4 to 3
>>> Migrating 1 to 4
>>> Migrating 3 to 4
>>> Migrating 0 to 3
>>> Migrating 2 to 3
>>
>> Wait.
>>
>> This may be non-optimal for cpusets, but maybe optimal migrate_pages,
>> especially
>> the usecase is HPC. I guess this is intended behavior. I think we need to hear
>> Christoph's intention.
>>
>> But, I'm not against this if he has no objection.
>
> The use case for this is if you have an app running on nodes 3,4,5 on your
> machine and now you want to shift it to 4,5,6. The expectation is that the
> location of the pages relative to the first node stay the same.
> Application may manage their locality given a range of nodes and each of
> the x .. x+n nodes has their particular purpose.
>
> If you justd copy 3 to 6 then the app may get confused when doing
> additional allocations since different types of information is now stored
> on the "first" node (which is now 4).

MPOL_INTERLEAVE is more simple situaltion. applications naturally assume the
memory is mapped intealeaving and application threads optimize for it. if we
broke intereaving, the applications may slow down.

--
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/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>