Cedric Le Goater wrote:
> Pavel Emelianov wrote:
>> OK. We have measured the nptl perf test for init namespace.
>> Summary - flat model is very light, Suka's patches break the 
>> kernel performance event when CONFIG_PID_NS is off.
>>
>>                  |    perf, s   | perf loss  |
>> -----------------+--------------+------------+
>> 2.6.22-rc4-mm2   | 13.79 ± 0.13 |    ---     |
>>                  |              |            |
>> suka  + PID_NS=n | 14.07 ± 0.13 |    2.0%    |
>> suka  + PID_NS=y | 14.06 ± 0.08 |    2.0%    |
>>                  |              |            |
>> pavel + PID_NS=n | 13.80 ± 0.07 |    0.0%    |
>> pavel + FLAT     | 13.80 ± 0.07 |    0.0%    |
>> pavel + MULTI    | 14.28 ± 0.08 |    3.5%    |
>> -----------------+--------------+------------+
> 
> is that on the same hardware you used last time ? 
> 2 * Intel(R) Xeon(TM) CPU 3.00GHz with 2GB RAM.

Yup.

>> I do believe that Suka's hierarchical model is better than mine,
>> but my point is: let's support the flat model as well.
> 
> OK. First thing we can do is to find what they have in common and 
> get that included. Then, after the first round, we might even find 
> some more to reach the MULTI model :) 

The [PREP xxx] series of patches does exactly this. It has the proc
changes, all the necessary things to work with pid numbers, all the
preparations in kernel/pid.c, signal handling, etc. Do you mind using
this?

The [MULTI xxx] series is just a demonstration of how this model 
can be done above my patches. I saw that Suka's model was faster
(and I think I know why) so I'm fine with throwing out my multilevel
model (only).

> That said, I'm perfectly fine with the FLAT model, because I think
> it covers nearly all the real world scenarii i know about : system 
> containers, application containers. 
> 
> C.
> 

_______________________________________________
Containers mailing list
Containers@lists.linux-foundation.org
https://lists.linux-foundation.org/mailman/listinfo/containers