[Wien] Blue Gene/L

[Peter Blaha]

Maybe I missed this previous posting.

Anyway: The current lapw1_mpi and lapw2_mpi will most likely NOT run 
well - maybe not at all . It has a large sequential overhead, both in 
terms of cpu-time, but even more important, in terms of memory. Thus for 
a "large case" (several hundreds of atoms) the memory/node would stay 
large, even when you could use 1000 processors.

We have (and still do) enhanced these mpi-versions significantly and 
most likely they will be released with the next update.

In lapw1_mpi, both, hamilt and hns scale very well with #cpus, both in 
time and memory.
The diagonalization goes with scalapack and is as good as scalapack is 
(I'd say this becomes the bottleneck, but maybe Blue/Gene has a better 
scalapack...?)
PS: We also have a greatly improved iterative diagonalization, but 
that's not in parallel yet.

lapw2 got an "additional" parallelization over the atoms and scales 
again nicely as long as the number of atoms is large (thus on 1000 
processors it may not scale perfectly anymore, but should be still quite 
fast.

Of course Blue/Gene makes sense only for cases with more than eg. 500 
atoms .... and don't expect to run anything "out of the box" on such 
machines....


Roberto Iglesias schrieb:
> Hi everybody!
> 
> Some time ago there was a post from Andrew Scott requesting information on Wien2k installations on an IBM Blue 
> Gene. Nobody answered and that's why I would like to ask again if anybody has some information on benchmarking 
> and scaling, as he requested, or, even better, on compiling and linking options on a Blue Gene/L machine.
> 
> Thanks in advance
> 
> Regards
> 
> Roberto
> 

-- 

                                       P.Blaha
--------------------------------------------------------------------------
Peter BLAHA, Inst.f. Materials Chemistry, TU Vienna, A-1060 Vienna
Phone: +43-1-58801-15671             FAX: +43-1-58801-15698
Email: blaha at theochem.tuwien.ac.at    WWW: 
http://info.tuwien.ac.at/theochem/
--------------------------------------------------------------------------