Re: Node affine NUMA scheduler

Jesse Barnes (jbarnes@sgi.com)
Thu, 14 Mar 2002 09:28:08 -0800


On Thu, Mar 14, 2002 at 03:54:12PM +0100, Erich Focht wrote:
> Jesse,
>
> thanks for running the tests. Actually "hackbench" is a bad example for
> the node affinity (though it's a good test for heavy scheduling). The code
> forks but doesn't exec and therefore all hackbench tasks have the same
> homenode. Also the tasks are not particularly memory bandwidth or latency
> hungry, therefore node affinity won't speed them up. I'm actually glad
> that they aren't slower, that shows that the additional overhead is small.

Alright, I'll try running some other numbers too, what can you
recommend other than aim and kernel compiles?

> Thanks for sending the macros for SGI_SN1/2, I'll include them. You
> probably use the DISCONTIGMEM patch, for that I append a small patch which
> "couples" DISCONTIGEMEM with the node affine scheduler such that pages
> will be allocated on the node current->node instead of the node on which
> the task is currently running. Hackbench might slow down a bit but
> AIM7 should improve.

Sounds good, I'll have to update those macros later too (Jack
reminded me that physical node numbers aren't always the same as
logical node numbers).

Jesse
-
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/