[Wien] Wien 2k benchmarks on Apple MacPRO (MacOS X Server 10.4.10)

Marcus Tegel tegel at cup.lmu.de
Fri Oct 12 16:47:47 CEST 2007


Hi!

I have compiled Wien 2k with ifort 10.0 / mkl (em64t) on my MacOS  
10.4.10 system (dual-Xenon 5300 8x3GHz).

I've also run a couple of different benchmarks with different amounts  
of tasks/threads. It just seems like the more jobs you run on the  
same system, the less important is the compiler and the more  
important the operating system's ability to handle multple tasks.  
MacOS 10.4 is known to be quite bad about this. Also the bandwidth to  
the RAM seems to be quite a bottleneck in my case.

For better comparison, I calculated the effective time per job for  
each benchmark (This time I took the total times from the benchmarks'  
output, which are given in mm:sec, averaged them and divided the  
average by the number of parallel tasks.)


#Tasks >	1		2		4		5		6		8
#threads v	113s	74s		53s		50s		52s		51s
1

2			91s		58s		55s		n/a		60s		60s

4			80s		64s		56s		n/a		63s		n/a

So it's obviously advantageous to run single-threaded jobs and at 4-8  
jobs the advantage of the INTEL compiler and MKL in comparison with  
GFORTRAN/LIBGOTO is only a couple percent.

Best wishes,
Marcus Tegel.


> Very interesting numbers, telling that a good compiler/library is  
> really
> important (and eventually worth some money).
>
> Still it would be interesting to have some multi-threading
> (probably mainly 2-threads) results, i.e.
> to be able to judge the "best possible performance" like comparing
> 2 jobs with 2 threads vs. 4 jobs with 1 thread,... or can one gain
> something with 4 jobs and 2 threads,...
>
>> My results were about 50% faster on a similar Mac Pro but running  
>> Fedora 7 and
>> using ifort 10.0 and mkl 9.1. I used the default compiler options  
>> plus -O3
>> -xT.
>>
>> 1 job with 1 thread:    115.459 sec
>> 2 jobs with 1 thread:  125.219 sec (62.610 sec/kpt)
>> 4 jobs with 1 thread:  167.083 sec (41.771 sec/kpt)
>> 6 jobs with 1 thread:  237.760 sec (39.627 sec/kpt)
>> 8 jobs with 1 thread:  311.175 sec (38.897 sec/kpt)
>>
>> On Oct 5, 2007, at 4:18 AM, Marcus Tegel wrote:
>>
>>> Hi!
>>>
>>> Benchmarks:
>>>
>>> MacPRO, dual-Xeon 5300, 8 x 3.0 GHz  182 sec gcc 4.01, gfortran,  
>>> libgoto (1
>>> job with 1 thread)
>>> MacPRO, dual-Xeon 5300, 8 x 3.0 GHz  254 sec (eq. 64 sec)    gcc  
>>> 4.01,
>>> gfortran, libgoto (4 jobs with 1 thread)
>>> MacPRO, dual-Xeon 5300, 8 x 3.0 GHz  337 sec (eq. 56 sec)    gcc  
>>> 4.01,
>>> gfortran, libgoto (6 jobs with 1 thread)
>>> MacPRO, dual-Xeon 5300, 8 x 3.0 GHz  448 sec (eq. 56 sec)    gcc  
>>> 4.01,
>>> gfortran, libgoto (8 jobs with 1 thread)
-------------- next part --------------
A non-text attachment was scrubbed...
Name: smime.p7s
Type: application/pkcs7-signature
Size: 2415 bytes
Desc: not available
Url : http://zeus.theochem.tuwien.ac.at/pipermail/wien/attachments/20071012/9ef94def/smime.bin


More information about the Wien mailing list