[gmx-developers] GROMACS 4.0.1 still broken on Bluegene

[Mark Abraham]

Erik Lindahl wrote:
> Hi,
> 
> Sorry, me too - I always use the default settings when retrieving my 
> to-do list, so I too missed it.
> 
> I'll see if I can make an instant 4.0.2 tomorrow, that also has 
> makefiles for windows. Then it's binary time!

Erik's helpfully fixed the earlier issues I've noted with BlueGene, 
along with some more I found today at 
http://bugzilla.gromacs.org/show_bug.cgi?id=250 and 
http://bugzilla.gromacs.org/show_bug.cgi?id=252

There can't have been any kind of real-world testing of this code:
* Any non-solvent charge groups in a neighbour-list of more than 6 atoms 
would have produced erroneous Coulomb energies and forces for a third of 
the interactions. (In my case, an equilibrated system blew up in under 
20 steps.)
* I suspect RF and cutoff have a sign error in generating fscal, but 
haven't tested or analysed the code to my satisfaction yet.
* Any double-precision build would have failed.

However, once I applied some fixes, a no-solvent-loop simulation of 50K 
atoms under PME is running stably on 32 processors under both single and 
double precision.

I haven't gotten around to looking at the water-optimized loops yet, but 
I will, and will test things out with gmxtest.pl as well.

Mark