[MMTK] Auto-discard notification

Sanzo Miyazawa miyazawa at smlab.sci.gunma-u.ac.jp
Wed Jun 21 12:16:01 CEST 2006


>From hinsen at cnrs-orleans.fr Tue Jun 20 00:59:41 2006
>>
>> I ran MMTK-2.5.13/Examples/MolecularDynamics/protein.py
>> with a modification to read a 1L2Y.pdb on several AMD
>> opteron servers.
>>
>> The energy values were different on different opteron computers.
>> ----------------------------------------------------------------

I printed each energy term and found that all except cosine_dyhedral_energy
are reproducible.


>>
>> I printed out the coordinates of atoms and
>> confirmed that they were the same.
>>
>> I tried to run it on several xeons and
>> confirmed that the energy values were exactly
>> the same between xeons.
>>
>> The energy values were the same on different xeon computers.
>> -------------------------------------------------------------

I found that different energy values for cosine_dyhedral_energy could be obtained 
even on xeons for longer chains.

>>
>> If someone has AMD chips, please try to run it,
>> and inform the result.
>> I need to know what is wrong in my systems.
>
>That looks strange... Could you please try to compile MMTK without  
>any compiler optimization?
>

I tried to use -O0 instead of -O3, but still
cosine_dyhedral_energy is not reproducible.

I inserted a debug line in bonded.c to print 4 atom indices, all atom positions, 
and dyhedral energy. 
The sequence of "4 atom indices", the reference order and
the way to specify the 4 atom indices,  were different depending on computers.

Is this normal?

If normal, then the order in summing energies can be different among computers,
and so the total sum may be slightly different, 
although the difference in the total cosine_dyhedral_energy 
is not as small as such a round-off error.

But I cannot imagine what kind of elements in the program cause such a result
depending on computers.

	Thanks for your help.





More information about the mmtk mailing list