On Wed, 19 Jan 2011, David Rientjes wrote:
> On Wed, 19 Jan 2011, Andi Kleen wrote:
>
> > cpusets didn't exist when I designed that. But the idea was that
> > the kernel has a first choice ("hit") and any other node is a "miss"
> > that may need investigation. So yes I would consider cpuset config as an
> > intention too and should be counted as hit/miss.
> >
>
> Ok, so there's no additional modification that needs to be made with the
> patch (other than perhaps some more descriptive documentation of a
> NUMA_HIT and NUMA_MISS). When the kernel passes all zones into the page
> allocator, it's relying on cpusets to reduce that zonelist to only
> allowable nodes by using ALLOC_CPUSET. If we can allocate from the first
> zone allowed by the cpuset, it will be treated as a hit; otherwise, it
> will be treated as a miss. That's better than treating everything as a
> miss when the cpuset doesn't include the first node.
To be more specific: It is the first zone of the zonelist that the cpuset
context provided for allocation from the node that the process is
currently executing on.
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxxx For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom policy in Canada: sign http://dissolvethecrtc.ca/
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>