Re: [PATCH 1/3 v2] slub: prefetch next freelist pointer in __slab_alloc()

[JoonSoo Kim]

2012/7/5 Pekka Enberg <penberg@xxxxxxxxxx>:
>> 2012/7/4 Pekka Enberg <penberg@xxxxxxxxxx>:
>>> Well, can you show improvement in any benchmark or workload?
>>> Prefetching is not always an obvious win and the reason we merged
>>> Eric's patch was that he was able to show an improvement in hackbench.
>
> On Wed, Jul 4, 2012 at 5:30 PM, JoonSoo Kim <js1304@xxxxxxxxx> wrote:
>> I thinks that this patch is perfectly same effect as Eric's patch, so
>> doesn't include benchmark result.
>> Eric's patch which add "prefetch instruction" in fastpath works for
>> second ~ last object of cpu slab.
>> This patch which add "prefetch instrunction" in slowpath works for
>> first object of cpu slab.
>
> Prefetching can also have negative effect on overall performance:
>
> http://lwn.net/Articles/444336/
>

Thanks for good article which is very helpful to me.

> That doesn't seem like that obvious win to me... Eric, Christoph?

Could you tell me how I test this patch more deeply, plz?
I am a kernel newbie and in the process of learning.
I doesn't know what I can do more for this.
I googling previous patch related to slub, some people use netperf.

Just do below is sufficient?
How is this test related to slub?

for in in `seq 1 32`
do
 netperf -H 192.168.0.8 -v 0 -l -100000 -t TCP_RR > /dev/null &
done
wait
--
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/