Re: [PATCH] dcache: Translating dentry into pathname without taking rename_lock

[Waiman Long <waiman.long@hp.com>]

On 09/04/2013 04:40 PM, John Stoffel wrote:
>>>>>> "Waiman" == Waiman Long<waiman.long@hp.com>  writes:
> Waiman>  In term of AIM7 performance, this patch has a performance boost of
> Waiman>  about 6-7% on top of Linus' lockref patch on a 8-socket 80-core DL980.
>
> Waiman>  User Range          |   10-100  | 200-10000 | 1100-2000 |
> Waiman>  Mean JPM w/o patch  | 4,365,114 | 7,211,115 | 6,964,439 |
> Waiman>  Mean JPM with patch | 3,872,850 | 7,655,958 | 7,422,598 |
> Waiman>  % Change            |  -11.3%   |   +6.2%   |   +6.6%   |
>
> This -11% impact is worisome to me, because at smaller numbers of
> users, I would still expect the performance to go up.  So why the big
> drop?
>
> Also, how is the impact of these changes on smaller 1 socket, 4 core
> systems?  Just because it helps a couple of big boxes, doesn't mean it
> won't hurt the more common small case.
>
> John

I don't believe the patch will make it slower with less user. It is more 
a result of run-to-run variation. The short workload typically completed 
in a very short time. In the 10-100 user range, the completion times 
range from 0.02-0.11s. With a higher user count, it needs several 
seconds to run and hence the results are more reliable.

Longman