Re: [PATCH] numa,sched: only consider less busy nodes as numa balancing destination

From: Rik van Riel
Date: Mon May 11 2015 - 10:44:14 EST


On 05/11/2015 08:44 AM, Jirka Hladky wrote:
> Hi Rik,
>
> we have results for SPECjbb2005 and Linpack&Stream benchmarks with
>
> 4.1.0-0.rc1.git0.1.el7.x86_64 (without patch)
> 4.1.0-0.rc2.git0.3.el7.x86_64 with your patch
> 4.1.0-0.rc2.git0.3.el7.x86_64 with your patch and AUTONUMA disabled
>
> The tests has been conducted on 3 different systems with 4 NUMA nodes
> and different versions of Intel processors and different amount of RAM.
>
>
> For SPECjbb benchmark we see
> -with your latest proposed patch applied
> * gains in range 7-15% !! for single instance SPECjbb (tested on
> variety of systems, biggest gains on brickland system, gains are growing
> with growing number of threads)

That is significant.

> * for multi-instance SPECjbb run (4 parallel jobs on 4 NUMA node
> system) on change in results
> * for linpack no change
> * for stream bench slight improvements (but very close to error margin)

Glad to hear the patch is not causing regressions.

Peter, can you queue up this patch in your sched tree, or
would you like me to make any changes to it first?

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