Re: [AMBER-Developers] AMBER Master Branch

From: B. Lachele Foley <lfoley.ccrc.uga.edu>
Date: Sat, 4 Jun 2016 17:52:51 +0000

Our queue was filled from the time this came out until just now. But, we finally tested, and do see a decent speed-up (~2x) in parallel, with a possible slight performance hit in serial. 149,500 atoms.

:-) Lachele

Dr. B. Lachele Foley
Associate Research Scientist
Complex Carbohydrate Research Center
The University of Georgia
Athens, GA USA
lfoley.uga.edu
http://glycam.org

________________________________________
From: Charles Lin <clin92.ucsd.edu>
Sent: Friday, June 3, 2016 3:15:30 PM
To: AMBER Developers Mailing List
Subject: Re: [AMBER-Developers] AMBER Master Branch

So there are performance enhancements for non-KNL systems primarily due to much better vectorization and hybrid MPI/OpenMP implementations.

KNL specific improvements are under the -mic2 switch. We also add in the SPDP version (experimental) which is similar to SPFP on cuda which also increases performance (uses -mic2_SPDP switch).

In particular for the latest Broadwell (36 core system) processors the proper input now will be:
PME: # of cores (36) & OMP_NUM_THREADS=2
GB: # of sockets (2) & OMP_NUM_THREADS=# of cores per socket (18)

Charlie
________________________________________
From: Hai Nguyen [nhai.qn.gmail.com]
Sent: Friday, June 03, 2016 12:11 PM
To: AMBER Developers Mailing List
Subject: Re: [AMBER-Developers] AMBER Master Branch

hi Benny,

I am confused here. Ross said "Performance changes here will be minimal for
pre V4 (Broadwell) hardware and most of the changes are focused on Knights
Landing Xeon Phi (to be released soon)."

Does this mean we need to wait and buy Knights Landing Xeon Phi to see the
difference? Or something else. thanks

Hai

On Fri, Jun 3, 2016 at 6:30 AM, Benny1 M <benny1.m.tcs.com> wrote:

> OpenMP does not come into play in PME on host and hence there is not much
> change in performance,
> or in the way the AMBER is run.
>
> Running PME:
> export I_MPI_PIN_MODE=pm
> export I_MPI_PIN_DOMAIN=auto
> mpirun -np NSOCKETS * NCORES_PER_SOCKET \
> $AMBERHOME/bin/pmemd.MPI
>
> On the other hand MPI-communication in GB workload has been reduced by
> using minimal MPI ranks and
> more OpenMP threads. This improves single node performance as well as
> helps in scaling across nodes.
> Improvement is seen in mid to large size workloads like nucleosome(~25K
> atoms) and rubisco( ~75K atoms).
> Try using IntelMPI + OpenMP and let us know if any issues are faced.
>
> Running GB:
> export I_MPI_PIN_MODE=pm
> export I_MPI_PIN_DOMAIN=auto
> mpirun -np NSOCKETS \
> -env OMP_NUM_THREADS=NCORES_PER_SOCKET*2 \
> -env KMP_AFFINITY="scatter,granularity=core" \
> -env KMP_STACKSIZE=10M \
> $AMBERHOME/bin/pmemd.MPI
>
>
> - Benny
>
>
>
> From: Jason Swails <jason.swails.gmail.com>
> To: AMBER Developers Mailing List <amber-developers.ambermd.org>
> Date: 03-06-2016 07:20
> Subject: Re: [AMBER-Developers] AMBER Master Branch
>
>
>
> It has something to do with OpenMP too, right? And then you have to be
> careful to count total threads as MPI*OMP threads to avoid thrashing. Do
> you only see the perf boost with OMP-MPI combo? That was my
> understanding...
>
> Is the exact recipe written down somewhere for how to take full advantage
> of this code? Because if Dave is having trouble using it "correctly", our
> users are highly unlikely to have better luck.
>
> --
> Jason M. Swails
>
> > On Jun 2, 2016, at 8:54 PM, Ross Walker <ross.rosswalker.co.uk> wrote:
> >
> > Hi Dave,
> >
> > Performance changes here will be minimal for pre V4 (Broadwell) hardware
> and most of the changes are focused on Knights Landing Xeon Phi (to be
> released soon).
> >
> > All the best
> > Ross
> >
> >> On Jun 2, 2016, at 16:59, David A Case <david.case.rutgers.edu> wrote:
> >>
> >> On Thu, Jun 02, 2016, Charles Lin wrote:
> >>>
> >>> So the Intel code has been in master for about a month now. We plan
> on
> >>> releasing the patch within a week.
> >>
> >> Still not sure when I am supposed to see speedups. I've tried
> pmemd.MPI
> >> runs (using Intel 16.0.3 compilers + MKL + mvapich2) for various
> systems, up to 64
> >> threads, and see no difference in speed for PME calculations. Have not
> >> tried any GB calculations.
> >>
> >> Is this only expected to speedup things when the -intelmpi (rather than
> -mpi)
> >> flag is set?
> >>
> >> Do you have specific examples of what systems one should expect
> speedups
> >> for?
> >>
> >> ...thx...dac
> >>
> >>
> >> _______________________________________________
> >> AMBER-Developers mailing list
> >> AMBER-Developers.ambermd.org
> >> http://lists.ambermd.org/mailman/listinfo/amber-developers
> >
> >
> > _______________________________________________
> > AMBER-Developers mailing list
> > AMBER-Developers.ambermd.org
> > http://lists.ambermd.org/mailman/listinfo/amber-developers
>
> _______________________________________________
> AMBER-Developers mailing list
> AMBER-Developers.ambermd.org
> http://lists.ambermd.org/mailman/listinfo/amber-developers
>
> =====-----=====-----=====
> Notice: The information contained in this e-mail
> message and/or attachments to it may contain
> confidential or privileged information. If you are
> not the intended recipient, any dissemination, use,
> review, distribution, printing or copying of the
> information contained in this e-mail message
> and/or attachments to it are strictly prohibited. If
> you have received this communication in error,
> please notify us by reply e-mail or telephone and
> immediately and permanently delete the message
> and any attachments. Thank you
>
>
> _______________________________________________
> AMBER-Developers mailing list
> AMBER-Developers.ambermd.org
> http://lists.ambermd.org/mailman/listinfo/amber-developers
>
_______________________________________________
AMBER-Developers mailing list
AMBER-Developers.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber-developers

_______________________________________________
AMBER-Developers mailing list
AMBER-Developers.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber-developers

_______________________________________________
AMBER-Developers mailing list
AMBER-Developers.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber-developers
Received on Sat Jun 04 2016 - 11:00:02 PDT
Custom Search