On Fri, Nov 22, 2013 at 10:59 AM, Ross Walker <ross.rosswalker.co.uk> wrote:
> Hi Yun,
>
> To expand a little bit on what Jason wrote. There are ways to tweak the QM
> performance a bit - you might be able to get a factor of 2 or 3
> improvement in performance but that's more or less it.
>
> Firstly: Compile with the Intel compiler and link to one of the latest
> versions of MKL.
>
> Second: set diag_routine=0 in the qmmm namelist. This will then try the
> various diagonalizers in MKL and select the fastest one. For QM atom
> counts above 40 or so the default diagonalizer is not the most efficient.
> - Timing info on each of the diagonalizers will be give in the mdout file.
>
> All the best
> Ross
>
> Hi Ross,
I also read the J Phys Chem A (2007) paper about SCC-DFTB in AMBER 9, which
states that "Extensive work is underway to make DFTB fully parallel in the
next AMBER release and it is anticipated that by that time the same DFTB
calculations as above will be no more than 2.0 times slower than the
corresponding all-classical simulation."
So I wonder if SCC-DFTB, also semi-empirical, would run faster than PM3 in
AMBER 12.
>
> On 11/22/13 10:32 AM, "Jason Swails" <jason.swails.gmail.com> wrote:
>
> >On Fri, 2013-11-22 at 10:19 -0800, yunshi11 . wrote:
> >> Dear all,
> >>
> >> I have a enzyme-substrate system that includes 80 QM+link atoms (with
> >>PM3,
> >> TIP3P water, PME implemented).
> >>
> >> As suggested in Amber12 manual, the scaling of such calculations are not
> >> good. In my test, both 24-core run and 12-core run gave me 20 ps per
> >>day,
> >> which means 50 days for 1 ns. Thus I wonder how to improve the
> >>performance.
> >
> >Stop using QM. As long as you are using QM on a part of your system,
> >there is a severe limit to how fast simulations will run. In fact, as
> >you have seen the best you can get is 20 ps/day, regardless of how many
> >CPUs you use.
> >
> >
> >> Should I adjust the number of cores according to the number of QM/link
> >> atoms? It seems kind of inefficient that NO QM atoms have been assigned
> >>to
> >> 4 of the 24 cores, while the other 20 cores take 4 QM atoms each.
> >
> >Semi-empirical QM is not particularly parallelizable, in stark contrast
> >to high-level ab-initio techniques. Unlike HF, DFT, and the more
> >expensive correlated methods, the bottleneck of semi-empirical
> >techniques is diagonalization of the Fock matrix rather than computing
> >the integrals. Diagonalization is not efficiently parallelized, which
> >means the slowest part of the QM calculation must be run by 1 CPU. You
> >can throw more CPUs at it, but they will all end up waiting for the
> >slowest CPU to finish diagonalizing the matrix.
> >
>
Does this mean that if I have some parallel computational power (100-200
CPU cores), I would get better performance when using HF or DFT (with
external QM programs rather than amber itself) for my QM region?
> >> How about using GPUs?
> >
> >No. First, Amber does not have GPU support for QM. Also, the matrix
> >diagonalization makes semi-empirical QM methods difficult to port to
> >GPUs efficiently, making it unlikely that Amber's QM will ever be ported
> >to GPUs.
> >
> >HTH,
> >Jason
> >
> >--
> >Jason M. Swails
> >BioMaPS,
> >Rutgers University
> >Postdoctoral Researcher
> >
> >
> >_______________________________________________
> >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 Nov 22 2013 - 12:00:04 PST
