[gmx-developers] Different risults on different platform

[Erik Lindahl]

Hi,

Molecular Dynamics is a chaotic process, so it is quite normal that you  
get different results on two different computers. Not only do they use  
different compiler optimizations, but the pentium version might even  
use SSE assembly loops.

There is of course always a possibility of a bug, but the "good" thing  
is that a bug in the MD code usually means the whole system blows up  
immediately. It is pretty rare that they just silently affect the  
results.

We do test Gromacs on a lot of systems, including Tru64, but it is  
simply impossible to test it with every single point release  
combination of the OS and compiler. There have been a couple (2 or 3,  
I'd say) of occasions when a new compiler version that previously  
worked has a bug that only affects high levels of optimization.

Start by testing simple thermodynamic properties like density and  
energy of the system. If these are similar, I'd say it is very unlikely  
there is a program or compiler bug on one platform. I would also  
recommend to use the latest version (3.1.4 right now), and you can try  
to disable the assembly loops on Pentium by setting the environment  
variable NOASSEMBLYLOOPS to anything (e.g. =1).

Cheers,

Erik




On Nov 26, 2003, at 10:35 AM, Pier Luigi Cristinziano wrote:

> This mail is addressed to gromacs developers who wrote the program. I  
> installed the 3-0-2 version on a Pentium IV cluster (4 PC) with Linux  
> RedHat O.S., and the 3-0-2 version on an HP ES45 (4 alpha processors)  
> with Thru64 O.S.. My simulations are on liquid crystals and I'm trying  
> to locate the isotropic transition point. I get very different risults  
> if I run the program starting from same point on both platform. To get  
> this point I have to execute several run at decreasing or increasing  
> temperature. The transition temperature on PC Linux cluster differs of  
> 50 degree from that on HP ES45. Could You explain me why happens this  
> and in any case which risults could I consider reliable? Do You tested  
> the program on all the platform for which gromacs is installable?
> Do You have some test case with output, so I could run it on my  
> machines and compare the output?
>                                                                         
>               Thank You
>                                                                         
>  Dr. Pier Luigi Cristinziano
>                                                                         
>  Department of Chemistry
>                                                                      
>       University of Basilicata
>                                                                         
>         Potenza - ITALY 
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: text/enriched
Size: 3093 bytes
Desc: not available
URL: <http://maillist.sys.kth.se/pipermail/gromacs.org_gmx-developers/attachments/20031126/7185c95b/attachment.bin>