> What is the sped you would get if you ran just ONE simple md, not remd?
>
> In simple md I am getting 4.43 ns/day on 16 CPU
I may have missed something in the first e-mail, but I believe it stated
you were running 64 replicas on 64 CPU (which I take to either mean 64
cores or 64 nodes). 64 replicas means 64 independent MD simulations.
If one simulation gets 4.43 ns/day on 16 CPUs (core or nodes), this is
equivalent to ~0.27 ns/day per CPU. Now running 64 you get ~0.11 ns/day
per CPU which means either communication is flooding the network or you
are overwhelming your I/O subsystem (disk) since you are writing 64x the
data. Since the exchanges were not often (15000 steps), but you are
dumping I/O every 1000 steps (now 64 times more often than the single
simulation) I think you are killing the disk. Are you writing to a
parallel file system? If not, dial back the output frequency to 5000
steps and see if this improves anything...
--tec3
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Wed Jun 07 2017 - 20:30:03 PDT
