On Tue, May 22, 2018, Ana Luisa Novo de Oliveira wrote:
>
> I´m currently applying for a supercomputer project and they asked me for
> the scability of the code in Marenostrum IV. Each node has two Intel
> Xeon Platinum chips, each of it with 24 processors. Do you have any data
> for systems running on this architecture? The minimal number of cores
> for these projects is 1024. My systems have around 175000 atoms. Would
> the amber code be suitable for my simulations?
Historically, Amber has been a poor choice for requirements like those
above, at least if you are considering fine-grained parallel runs, where
a single trajectory is divided up among 1024 processors. Until very
recently, Amber would scale to at best ~100-150 cores on a single
trajectory.
People can (and do) use Amber on large numbers of cores to carry out
replica exchange calculations, where many trajectories are being run at
once, and each single trajectory only uses a small number of cores,
usually on a single node.
There has been some recent work on getting better fine-grained parallel
performance: http://ambermd.org/intel/midpoint.htm. But this effort is
still listed as a "prototype", and is primarily aimed at Xeon Phi, at
least as far as I understand. (Maybe others on the list can give more
detailed information.)
I hope this helps, and I hope others who have more direct experience
that I do with chime in.
...dac
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Tue May 22 2018 - 14:00:02 PDT
