[gmx-users] CVS support for threads [was: qm_mopac.c problems in building of 3.3.1]

[Erik Lindahl]

Hi,

On Oct 9, 2006, at 5:38 PM, Anton Feenstra wrote:

>
> Nevertheless, on newer architectures, threads would ultimately be  
> the better choice. They're working on it and according to David van  
> der Spoel it will be in 4.0 (that was on the Developer's list in  
> May '04... Brrrrr, as David would put it ;-).

Our main problem is that "SMP architectures" has changed a _lot_ the  
last 2 years. Instead of having a pool of CPUs reading the same  
memory we now need to be careful to avoid NUMA issues, i.e. each  
processor "owns" a part of memory, and it is really bad to just  
access it randomly (read: worse than our current MPI parallelization,  
and thus rather pointless).

We've completely overhauled the MPI parallelization the last year and  
now have absolutely amazing scaling in CVS. All nonbonded kernels are  
already 100% multithread-capable, but I'm still having some problems  
with the neighborsearching, which is the last (big) bit in the puzzle.

Suffice to say, we will get there, but it won't be enabled in the  
release until it is useful ;-)

Cheers,

Erik