On Mon, Sep 21, 2015, Sabyasachi Sahoo wrote:
>
> I am a parallel programming researcher and would like to contribute to
> AMBER molecular dynamics software by helping to nail down any bottlenecks
> that occur in scaling of the software on multiple CPUs (and probably
> improved performance in exa-scale era). I am already in process of
> collecting profiling results to identify the phases that can be improved.
One (very non-obvious) web page to look at is this one:
http://ambermd.org/namd/namd_amber.html
This shows scaling behavior of NAMD vs. pmemd, and shows the poor scaling of
pmemd for more than 100-200 threads. Note that:
* As with most MD codes, pmemd would probably scale better for larger
systems
* Neither NAMD (at least version 2.7) nor pmemd are really ready for
exa-scale computing, absent the use of meta-procedures like replica
exchange
The general feeling among Amber developers (and others can chime in here,
especially if they disagree!) is that more than tweaks would be required
to significantly improve pmemd scaling. There are some fundamental choices
about data structures and algorithms that inhibit fine-grained parallel
scaling. Efforts at radical surgery should take account of lessons learned
from GPUs (e.g. JCTC 9:3878, 2013), and be cognizant of the "competition" from
NAMD, Gromacs, Desmond, OpenMM, CHARMM (with the "domdec" option), etc.
That doesn't mean you shouldn't try! Better information about scaling
bottlenecks might suggest some novel solutions that we haven't considered
before.
....dac
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Mon Sep 21 2015 - 12:00:03 PDT
