LAMMPS WWW Site - LAMMPS Documentation - LAMMPS Mailing List Archives
Re: [lammps-users] How to accelerate the calculation speed by using existing hardware ?
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [lammps-users] How to accelerate the calculation speed by using existing hardware ?


From: 王淑婷 <wangshuting@...6572...>
Date: Tue, 11 Jul 2017 11:09:10 +0800 (GMT+08:00)

Dear lammps-users, 

Thanks a lot to Axel Kohlmeyer. It's very kind of you to reply me. I agree with your explanation about slow calculation speed caused by hybrid pair potential. Your suggestion about how to accelerate the calculation speed is very useful.

I have a new question now. I use the command "mpirun -n 1 lmp_mpi -sf gpu -pk gpu 1 < in.W " to calculate the thermal conductivity. But it indicates the ERROR : Cannot open EAM potential file eam/alloy. Here is my input script blew. Could anyone give me some suggestions? I don't know what's wrong with my input script. Thank you for your help in advance.

Best wishes.
Shuting Wang

############################################
#-----------------------------Initialization

variable                      L equal 200/4.07
variable                      W equal 30/4.07
variable                      H_0 equal 50/4.07
variable                      H_1 equal 20/4.07
variable                      H_2 equal 30/4.07
variable HOT_R equal round($L*0.125)
variable HOT_L equal ${HOT_R}-1
variable COLD_L equal round($L*0.875)
variable COLD_R equal ${COLD_L}+1

units                metal
dimension                     3
boundary                      p p p
atom_style                    atomic

#------------------------------Atom Defination

lattice                       fcc 3.157 
region                        W block 0 $L 0 $W ${H_1} ${H_2}
region                        total block 0 $L 0 $W 0 ${H_0}
create_box                    1 total
create_atoms                  1 region W

region hot block ${HOT_L} ${HOT_R} 0 $W 0 ${H_0}
region cold block ${COLD_L} ${COLD_R} 0 $W 0 ${H_0}
group hot region hot
group cold region cold

mass                          1 196.96654

#Force-field parameters
#B=1,N=1

pair_style      eam/alloy
pair_coeff      * * eam/alloy W.eam.alloy

neighbor                      2.0 nsq
neigh_modify                  delay 0 every 10

#---------------------------------Define Setting

variable                      ke equal ke
variable                      pe equal pe
variable                      press equal press
variable                      vol equal vol
variable                      etotal equal etotal
variable                      temp equal temp
compute                       tot_temp all temp
compute                       myKE all ke/atom
variable                      temp1 atom c_myKE/0.0001292355
compute                       hot_temp all temp/region hot
compute                       cold_temp all temp/region cold

min_style cg
minimize 1.0e-4 1.0e-6 100 1000
reset_timestep 0


> -----原始邮件-----
> 发件人: "Axel Kohlmeyer" <akohlmey@...24...>
> 发送时间: 2017年7月10日 星期一
> 收件人: "王淑婷" <wangshuting@...6572...>
> 抄送: "LAMMPS Users Mailing List" <lammps-users@lists.sourceforge.net>
> 主题: Re: [lammps-users] How to accelerate the calculation speed by using existing hardware ?

> On Sun, Jul 9, 2017 at 10:40 PM, 王淑婷 <wangshuting@...6572...> wrote:
> > Dear lammps-users,
> >
> > Now I have several intel xeon processors e5 v4 with 280 cores. I find that
> > the calculation speed is the fastest by using 54 cores when I use hybrid
> > force field.
> >
> > I still have a NVIDIA Tesla P100 GPU. When the force field is tersoff, the
> > calculation speed tested by one GPU core and one CPU core is four times
> > higher than the 28 CPU cores. But when the force field is hybrid, the
> > calculation speed tested by one GPU and one CPU core is two times lower than
> > the 28 CPU cores.

> that is because with a hybrid pair potential, you must compute the
> neighbor lists on the CPU and then transfer it to the GPU, even if all
> individual pair styles are GPU accelerated.
> that will cause a significant slowdown, in your case, since you are
> using only one CPU. you can try to minimize this by using multiple
> CPUs with the same GPU, but the additional overhead of computing
> neighbor lists on the CPU and transferring significant additional data
> to the GPU cannot be completely avoided.

> > I don't know how to use existing hardware to accelerate the calculation
> > speed when my force field is hybrid. It's OK if I should buy some more
> > hardware. Could you please give me some suggestions? Thank you for your help
> > in advance.

> the best way to get the most performance with the GPU package, is to
> avoid a hybrid pair style.
> there are also many other, simulation specific factors, that can
> affect performance, where it is not possible to give general advice.

> axel.

> >
> > Yours sincerely
> > Shuting Wang
> >
> >
> >
> >
> >
> >
> > ------------------------------------------------------------------------------
> > Check out the vibrant tech community on one of the world's most
> > engaging tech sites, Slashdot.org! http://sdm.link/slashdot
> > _______________________________________________
> > lammps-users mailing list
> > lammps-users@lists.sourceforge.net
> > https://lists.sourceforge.net/lists/listinfo/lammps-users
> >



> -- 
> Dr. Axel Kohlmeyer  akohlmey@...24...  http://goo.gl/1wk0
> College of Science & Technology, Temple University, Philadelphia PA, USA
> International Centre for Theoretical Physics, Trieste. Italy.