AMBER Archive (2008)

Subject: AMBER: Fw: Amber9 vs Amber10+MKL benchmark

From: Francesco Pietra (chiendarret_at_yahoo.com)
Date: Sat May 24 2008 - 08:51:53 CDT

Sorry, I forgot to mention that both MD simulations were with 8 CPUs, NUMA-type machine.

--- On Sat, 5/24/08, Francesco Pietra <chiendarret_at_yahoo.com> wrote:

> From: Francesco Pietra <chiendarret_at_yahoo.com>
> Subject: Amber9 vs Amber10+MKL benchmark
> To: "Amber" <amber_at_scripps.edu>
> Date: Saturday, May 24, 2008, 6:48 AM
> I am comparing two portions of the same classical MD
> production (pore protein with a large ligand in a POPC
> hydrated membrane) for Amber9 vs Amber10 on the same
> hardware. As I found Amber 9 slower by some 5% (evaluated
> from either the "Total time" or the "Real
> time"), I am presenting the situation.
>
> The in file reads:
>
> prod protein ligand POPC membrane box 80x80
> &cntrl
> imin=0, irest=1, ntx=5,
> nstlim=333334, dt=0.0015,
> cut=10, ntb=2, ntp=1, taup=2.0,
> ntc=2, ntf=2,
> ntpr=1000, ntwx=1000,
> ntt=3, gamma_ln=2.0,
> temp0=300.0,
> /
>
>
> Portion 4: Amber 10 compiled with ifor+MKL 10.1.015,
> support openmpi-1.2.6 (compiled same Intel).
>
>
> Portion 3: with Amber 9 compiled with ifort 9.1.036,
> support openmpi-1.2.3 (compiled same Intel).
>
> I expected little gain from MKL for classical MD because of
> the overhead, though I was surprised that Amber 10 has
> suffered so much to turn out to be slower.
>
> That the MKL are installed correctly is proved by a 40%
> speed gain by using MKL in a heavy docking simulation (were
> ca 8GB MEM are used).
>
> I assume that each nstlim=333334 run of MD should require
> the same time of execution. If this is not wrong, is the
> situation presented amenable to a comment?
>
> Thanks
> francesco pietra

      
-----------------------------------------------------------------------
The AMBER Mail Reflector
To post, send mail to amber_at_scripps.edu
To unsubscribe, send "unsubscribe amber" (in the *body* of the email)
      to majordomo_at_scripps.edu