[Devel] Re: nptl perf bench and profiling with pidns patchsets

Pavel Emelianov xemul at sw.ru
Mon Jun 4 07:31:04 PDT 2007


Cedric Le Goater wrote:
> Pavel Emelianov wrote:
>> Serge E. Hallyn wrote:
>>> Quoting Kirill Korotaev (dev at sw.ru):
>>>> Cedric,
>>>>
>>>> just a small note.
>>>> imho it is not correct to check performance with enabled debug in memory allocator
>>>> since it can influence cache efficiency much.
>>>> In you case looks like you have DEBUG_SLAB enabled.
>>> Hm, good point.  Cedric, did you ever run any tests with profiling and
>>> debugging turned off?
>> I'd like to add that the results-for-comparison have to be run
>> with profiler turned off. Further, if we need to know what the
>> bottleneck is, the profiler is on, but the numbers get are not
>> trusted.
>>
>> Cedric, may I ask you to rerun the tests with both the debug and
>> the profiler turned off and report the results again?
> 
> sure. let me do all debug=off first because i'm interested in some
> figures.

Just to be sure. When I tested the namespaces I made the node
clean from any daemon that could spoil the results and made the
cache hot for the files involved in testing. Otherwise the results
could have more than 5% of accuracy which is not enough...

> so what do you think of the nptl perf benchmark to evaluate our 
> progress ? 

If this is just a spawn test for threads, then I think this is
not enough. This test *is* important, but we have to check some
more issues when talking about the namespaces.

I will look at this test closer tomorrow for more competent answer.

> C.

Thanks,
Pavel


More information about the Containers mailing list