[pooma-dev] Re: [PATCH] don't bench too much for boundschecking

[Jeffrey D. Oldham]

Richard Guenther wrote:

> Jeffrey D. Oldham wrote:
>
>> Richard Guenther wrote:
>>
>>> This reduces particle benchmarking if POOMA_BOUNDS_CHECK is on
>>> to one time with 100 particles (it takes an awful lot of time).
>>>
>>> Ok?
>>>
>>> Richard.
>>>
>>>
>>> 2004Aug27  Richard Guenther <richard.guenther at uni-tuebingen.de>
>>>
>>>     * src/Particles/tests/particle_tests.h: for POOMA_BOUNDS_CHECK
>>>     reduce default problem size(s).
>>>  
>>>
>>> ------------------------------------------------------------------------ 
>>>
>>>
>>> Index: Particles/tests/particle_tests.h
>>> ===================================================================
>>> RCS file: 
>>> /home/pooma/Repository/r2/src/Particles/tests/particle_tests.h,v
>>> retrieving revision 1.22
>>> diff -u -u -r1.22 particle_tests.h
>>> --- Particles/tests/particle_tests.h    23 Aug 2004 18:44:17 
>>> -0000    1.22
>>> +++ Particles/tests/particle_tests.h    27 Aug 2004 15:45:38 -0000
>>> @@ -400,8 +400,13 @@
>>>   // Default parameters for the benchmark.
>>>
>>>   int iters = 1000;
>>> +#if POOMA_BOUNDS_CHECK
>>> +  int startnumparticles = 100;
>>> +  int endnumparticles = 100;
>>> +#else
>>>   int startnumparticles = 100;
>>>   int endnumparticles = 10000;
>>> +#endif
>>>   int multnumparticles = 10;
>>>   double movefrac = 0.1;
>>>   bool usesync = false;
>>>  
>>>
>> I appreciate the desire to reduce the running time when bounds 
>> checking occurs, but this new set of values ensures the loop runs 
>> only once.  Without the change, the loop runs three times.  It would 
>> be nice to have the loop run at least twice to ensure it does not 
>> break something when the loop counter is updated.  To do this, set 
>> endnumparticles to 1000 when bounds are checked.  A more complex 
>> change is to change 'endnumparticles' to 'startnumparticles * 
>> multnumparticles' and reducing 'multnumparticles' to a smaller value 
>> greater than one.
>
>
> Another possibility would be to reduce iters to 10 and endnumparticles 
> to 1000.  Would this be ok?
>
> Thanks,
> Richard.
>
I would prefer to minimize the differences so I would prefer
a) Reducing iters from 1000 to 10 and leaving startnumparticles and 
endnumparticles unchanged
over
b) Reducing iters from 1000 to 10 and reducing endnumparticles to 1000,
but I am not interactively running the tests so I do not know how long 
they require.

Choose whichever choice yields the most reasonable testing time, e.g., 
1-5 minutes maximum.

Thanks for catching this.

-- 
Jeffrey D. Oldham
oldham at codesourcery.com