lkml.org 
[lkml]   [2007]   [Jan]   [17]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
SubjectRe: [RFC 5/8] Make writeout during reclaim cpuset aware
Date
On Wednesday 17 January 2007 15:36, Paul Jackson wrote:
> > With a per node dirty limit ...
>
> What would this mean?
>
> Lets say we have a simple machine with 4 nodes, cpusets disabled.

There can be always NUMA policy without cpusets for once.

> Lets say all tasks are allowed to use all nodes, no set_mempolicy
> either.

Ok.

> If a task happens to fill up 80% of one node with dirty pages, but
> we have no dirty pages yet on other nodes, and we have a dirty ratio
> of 40%, then do we throttle that task's writes?

Yes we should actually. Every node should be able to supply
memory (unless extreme circumstances like mlock) and that much dirty
memory on a node will make that hard.

> I am surprised you are asking for this, Andi. I would have thought
> that on no-cpuset systems, the system wide throttling served your
> needs fine.

No actually people are fairly unhappy when one node is filled with
file data and then they don't get local memory from it anymore.
I get regular complaints about that for Opteron.

Dirty limit wouldn't be a full solution, but a good step.

> If not, then I can only guess that is because NUMA
> mempolicy constraints on allowed nodes are causing the same dirty page
> problems as cpuset constrained systems -- is that your concern?

That is another concern. I haven't checked recently, but it used
to be fairly simple to put a system to its knees by oversubscribing
a single node with a strict memory policy. Fixing that would be good.

-Andi
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

\
 
 \ /
  Last update: 2007-01-17 07:25    [W:0.137 / U:0.696 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site