Re: [Qemu-devel] [PATCH for-1.5 0/9] Disable expensive QOM cast debugging for official releases

[Anthony Liguori <aliguori@us.ibm.com>]

Aurelien Jarno <aurelien@aurel32.net> writes:

> On Fri, May 10, 2013 at 02:16:34PM +0200, Paolo Bonzini wrote:
>> Cast debugging can have a substantial cost (20% or more, measured by
>> Aurelien on qemu-system-ppc64).  Instead of adding special-cased "fast
>> casts" in the hot paths, we can just disable it in releases.  At the
>> same time, add tracing facilities that simplify the analysys of those
>> problems that cast debugging would reveal.
>> 
>> At least patches 1-7 are for 1.5.
>> 
>> Paolo Bonzini (9):
>>   qom: improve documentation of cast functions
>>   qom: allow casting of a NULL class
>>   qom: add a fast path to object_class_dynamic_cast
>>   qom: pass file/line/function to asserting casts
>>   qom: trace asserting casts
>>   qom: allow turning cast debugging off
>>   build: disable QOM cast debugging for official releases
>>   qom: simplify object_class_dynamic_cast, part 1
>>   qom: simplify object_class_dynamic_cast, part 2
>> 
>>  configure            | 20 ++++++++------
>>  include/qom/object.h | 40 ++++++++++++++++++++++-----
>>  qom/object.c         | 77 ++++++++++++++++++++++++++++++++++------------------
>>  trace-events         |  3 ++
>>  4 files changed, 99 insertions(+), 41 deletions(-)
>> 
>
> I have tested this series with qemu-system-ppc64, on a Core i7 2600 CPU.
> The process was set to a single core using taskset. Inside the guest
> (Debian ppc64 from debian-ports), I ran the command three times:
>
>   lintian g++-4.8_4.8.0-6_ppc64.deb
>
> I used lintian as it's a perl code, that trigger the discussion about
> sparc/ppc comparison.
>
> First of all with this patch series, the object_class_dynamic_cast calls
> went down to below 0.1% when using perf top. Before the patch series was
> applied, the command took in average on 3 runs 142.4s. With the patch
> series, it went down to 129.8s, so almost 9% faster.

I just posted another patch which I believe will also reduce this
overhead without eliminating the checks.

We do a staggering number of casts...  The patch I posted makes the
overwhelming majority of them nothing more than a single pointer
comparison and a couple derefs.

Regards,

Anthony Liguori

>
> To improve the performance a bit more, and come back to the same kind of
> code as before, we should move simple accessors from qom/*.c to
> include/qom/*.h and mark them as inline, so that they can be removed by
> the compiler. Currently, even if the function is simple it's still a
> call/ret in the hot path instead of a simple pointer addition.

>
> -- 
> Aurelien Jarno	                        GPG: 1024D/F1BCDB73
> aurelien@aurel32.net                 http://www.aurel32.net