On Fri, Jun 25, 2021, Scott Le Grand wrote:
>Can you get more specific than RISM and XRay stuff? Is it because of better
>OpenMP support?
I was comparing ifort vs gfortran10 on code that makes heavy use of array
operations. Biggest win was in using OpenMP for 3D-RISM: in one example
gfortran took 701 sec., ifort 317 sec.
In other applications, I can get speedups by using exploting
hyperthreading, e.g. asking for 80 threads on a node that has 40 physical
cores is considerably faster than requesting just 40 OMP threads.
I'm not trying to push people to Intel compilers, and the above examples are
not representative of "typical" Amber jobs. I'm quite busy right now, but
if I get bored, I might try some more systematic benchmarks. I just wanted
to point out the relatively recent change in availability. It can be
installed anywhere, i.e. without root privileges (although I haven't tried
this.) Just getting a different set of compiler warnings might justify the
setup effort.
....dac
_______________________________________________
AMBER-Developers mailing list
AMBER-Developers.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber-developers
Received on Fri Jun 25 2021 - 18:30:03 PDT
