Re: QtHandles and performance

[Michael Goffioul]

On Fri, Jan 10, 2014 at 4:49 PM, John W. Eaton <address@hidden> wrote:

On 01/09/2014 04:37 PM, John W. Eaton wrote:

On 01/09/2014 04:30 PM, Rik wrote:

On 01/09/2014 12:06 PM, octave-maintainers-request@octave.org wrote:

Do we need to re-measure the performance penalty before we can make an
informed decision, or are we confident in the 10% number?

Measuring again is probably a good idea.  I compiled with optimization
enabled (default -O2 for GCC) and timed running the test suite to compare.

OK, I did that test again and this is what I see (default -g -O2
compiler options for both cases):

  without atomic refcount (current default):

    make[1]: Leaving directory `/scratch/jwe/build/octave-opt/test'
    401.45user 49.45system 7:50.31elapsed 95%CPU (0avgtext+0avgdata 415228maxresident)k
    153352inputs+20904outputs (172major+16480393minor)pagefaults 0swaps

  with atomic refcount:

    make[1]: Leaving directory `/scratch/jwe/build/octave-opt-atomic/test'
    405.32user 49.07system 7:51.65elapsed 96%CPU (0avgtext+0avgdata 415348maxresident)k
    85768inputs+20920outputs (54major+16128734minor)pagefaults 0swaps

That's about, like, no difference.  Hmm.  And yes, octave-opt/config.h
has

  /* #undef USE_ATOMIC_REFCOUNT */

and octave-opt-atomic/config.h has

  #define USE_ATOMIC_REFCOUNT 1

Well, that's not the result I was expecting.  So I guess there's no
reason to NOT use atomic refcount.  I ran the test with GCC 4.8.2 on an
x86_64 system.  Maybe it was worse with earlier versions of GCC?  I
guess I could try that just to see if I can confirm what I remember as
the previous result.  Or we could also look up the previous discussion
in the mailing list archives.

Results could be different on another architecture, like x86, as atomic referencing is eventually implemented using special ASM instructions.

Michael.