On Wed, Sep 11, 2013 at 4:34 AM, Hai Nguyen <nhai.qn.gmail.com> wrote:
> On Wed, Sep 11, 2013 at 2:03 AM, Francesco Pietra <chiendarret.gmail.com>
> wrote:
> > Hello Hai:
> > Can a debug trial be carried out on a shared memory GPU machine (the one
> I
> > use currently) or are multiple physical nodes required? With NAMD,
> physical
> > nodes are needed.
>
MPI allows for distributed computing because each thread is given its own
chunk of memory, with none shared between the different threads. [1] While
this allows MPI to be spread across multiple nodes, there is certainly no
rule against providing separate memory to multiple threads from the same
source of RAM.
As far as Amber (or any MPI program) is concerned, there is no difference
between running 8 threads on 1 node with shared memory or 8 threads on 8
nodes with no shared memory.
I'm not familiar with REMD in NAMD. I know NAMD uses charm++, which is a
different parallelization scheme (although it can optionally be compiled
using MPI as its message-passing back-end, I believe), and it also takes
advantage of shared memory parallelism, if available.
Hopefully this answers your question,
Jason
--
Jason M. Swails
BioMaPS,
Rutgers University
Postdoctoral Researcher
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Wed Sep 11 2013 - 05:00:04 PDT
