This is certainly not my field... then I may be wrong, but I suspect he should be using an AMD EPYC for running with 64 threads...
Any way... as far as I understand, such architecture was build for servers and uses Non-uniform memory access (NUMA) [
https://en.wikipedia.org/wiki/Non-uniform_memory_access] to scale up the number of processors, which means not all processors are able to directly access the memory... indeed, there are lots of gaming blogs out there explaining why such chip doesn't performs as expected on games... its strength is on several individual tasks not on sheared ones... So it may not be rare the lost of performance after 32 threads...
Best,
Matias
------------------------------------
PhD.
Researcher at Biomolecular Simulations Lab.
Institut Pasteur de Montevideo | Uruguay
[
http://pasteur.uy/en/laboratorios-eng/lsbm]
[
http://www.sirahff.com]
----- Mensaje original -----
De: "Daniel Roe" <daniel.r.roe.gmail.com>
Para: "AMBER Mailing List" <amber.ambermd.org>
Enviados: Viernes, 21 de Septiembre 2018 14:18:41
Asunto: Re: [AMBER] Running GIST calculation in parallel
Hi,
Since based on your timings you stop getting a speedup beyond 32
threads, I recommend using a max of 32 threads. :-)
Assuming you have the Ryzen 32-core processor this makes sense -
cpptraj does not benefit from hyperthreading (I think AMD calls it
simultaneous multi-threading but it's essentially the same thing.
-Dan
On Fri, Sep 21, 2018 at 10:57 AM M RCC <mkr3j2c1.gmail.com> wrote:
>
> Thank you Daniel for your suggestion.
> For benchmarking I did the following
> 100 frames [1 out of 6 boxes]
> 4 threads: 218 seconds
> 8 threads: 161 seconds
> 10 threads: 143 seconds
> 12 threads: 133 seconds
> 14 threads: 133 seconds
> 16 threads: 126 seconds
> 32 threads: 118 seconds
> 64 threads: 136 seconds
>
> Am splitting the boxes for calculation. so i think i dont need to worry
> about memory, though i have 120GB of RAM.
>
> Among these in which one i can expect faster performance?
>
> Thanks in advance
>
>
>
>
> On Thu, Sep 20, 2018 at 3:27 PM M RCC <mkr3j2c1.gmail.com> wrote:
>
> >
> > Dear Amber Users,
> >
> > I am trying to run GIST calculation in parallel with 64 core AMD system
> > using Ambertools18.
> > when am using "mpirun -np 64 cpptraj.MPI -i gist.in" , i got error with
> > core dump.
> >
> > then i am using "cpptraj.MPI -i gist.in" (even assigning
> > OMP_NUM_THREADS=64) program is running in only one core.
> >
> > How i can utilise all 64 core for GIST calculation?
> >
> >
> _______________________________________________
> AMBER mailing list
> AMBER.ambermd.org
> http://lists.ambermd.org/mailman/listinfo/amber
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Fri Sep 21 2018 - 11:30:02 PDT
