[gmx-users] Scaling on superclusters

[Florian Haberl]

Hi,

> > has anyone been tried to run gromacs on some more cpus, let me say > 100
> > cpus up to 1000  ( for example on a sgi altix, Itanium-2 with numalink).
> > Will there be still scaling or will there be a limit by  the non parallel
> > part of gromacs, and where will there be this limit?
>
> I will try next week. I'm not sure people have really tried with more
> than 32 CPUs. 32 works for large problems (e.g. 100,000 atoms) and
> without PME (although a fix for that is underway).
>
We got here a 32 cpu (16 x dual xeon2 emt64 with 3,2 ghz with iband 
interconnection) and i^ll simulate 30 k atoms (406 aa) with PME and it^s 
scales rather good (speedup 2cpus /32 cpus= 1: around 13 best would be 16 so 
really nice, and i got around 13 GFlops ). Soon we will get 24 iband so that 
would be around 48 cpus ...


Greetings,

 Florian

-- 
-------------------------------------------------------------------------------
 Florian Haberl                            Universitaet Erlangen/
 Computer-Chemie-Centrum    Nuernberg
                                                      Naegelsbachstr. 25
                                                      D-91052 Erlangen
  Mailto: florian.haberl AT chemie.uni-erlangen.de
 -------------------------------------------------------------------------------