[gmx-users] Athlon Instabilities

[Stefano Piana]

Hallo GMX users!
we are running a small dual athlon cluster. However we have noticed that the 
machines were very unstable with several programs. In particular we found out 
that gromacs crashes at random places relatively often (as reported by other 
users in February I guess...) with not reproducible settle failure messages.

We noticed that something similar happens with other programs too (Turbomole, 
Sander).
After some investigations we found out that the problem with Amber and 
Turbomole was apparently due to the usage of -Mvect=sse within the PGF 
compiler. In particular when the -Mvect=sse flag is used without 
-Mvect=cache_align the numbers produced by the program are no longer the same 
when the same run is repeated twice and eventually the calculation gets unstable.

  In gromacs I have tried to recompile without the assembly inner loops, but 
apparently, even with no assembler inner loops, when I run 10 times the same 
calculation I obtain 10 different results with differences sometimes of the 
order of 0.1-1%.

Now the question is:
I am not too familiar with the gromacs code; could it be that even if I 
configure with the --enable-fortran --disable-x86-asmc flags, SSE instructions 
are still used by the program?
Is it normal that gromacs produces slightly different results (0.1-1%) already 
after 100 steps of MD when the same calculations is run twice?
Could it be simply an heating problem that comes out more often when sse 
instructions are called?
Did anybody else experienced anything similar?
Thanks for any suggestion.

Stefano
-- 

ETHZ Honggerberg - HCI H203
8093 Zurich, Switzerland
Tel: +41 1 6323054
Fax: +41 1 6321090
piana at inorg.chem.ethz.ch