(Paul can you post this to the list)
I am using Lammps 2001. By any chance is there some documentation for
LAMMPS Output (Output to Screen while the simulation is running). I can
only see these items in the Lammps Website:
This doc page for the current LAMMPS may be helpful, since the output is
not that different for 2001 vs 2005.
http://www.cs.sandia.gov/~sjplimp/lammps/doc/Section_start.html#2_5
If not, can you be more specific about what screen output doesn't make
sense?
Steve
Jan-Michael,
Comparison of �Total time� is probably the most
important. For a system like yours, the long-range
electrostatics will likely be a large part of the
computational expense. PPPM =
particle-particle/particle-mesh is very similar to,
but slightly better accuracy per cost than PME =
particle mesh Ewald. PPPM is the mesh-based long-range
electrostatics solver in LAMMPS. Both long-range
electrostatics methods work well and are typically
much less expensive than traditional Ewald. With that
said, your problem (highly charged, low particle
density, right?) may be a case where traditional Ewald
would be less expensive. You might consider trying a
large real-space cutoff (approaching L/2) and maybe
even traditional Ewald instead of PPPM to see if you
get better speed.
Your �PPPM timing info� probably looks something like
the following from the old fortran version of LAMMPS:
PPPM timing info:
Make_rho time = 192.022138710533
Make_rho of long time = 4\.12816982866073
Poisson time = 3245\.90243224019
Poisson of long time = 69.7817271359004
Electric_field time = 209.175319532553
Electric_field of long time = 4\.49693586788172
Brick2fft time per timestep =
6\.790304183959961E\-003
Total brick2fft time = 339\.515209197998
Brick2fft of long time = 7.29903568616867
Fillbrick time per timestep =
7.315695285797119E-003
Total fillbrick time = 365.784764289856
Fillbrick of long time = 7\.86378923735178
FFT time per timestep = 9\.650218486785889E\-002
Total FFT time = 4825\.10924339294
FFT of long time = 103.732155741655
Setup time per timestep = 1.942002773284912E-003
Total setup time = 97.1001386642456
Setup % of other time = 2.94728361346330
I won�t spell out all of the details of what each of
these categories mean unless you really want me to do
so. You could look at pppm.f to see exactly what the
code is doing in each of these categories. To make a
long story short, these items are just the steps that
LAMMPS has to go through to properly set up and
execute the PPPM long-range electrostatics. Similarly
LAMMPS does a complete breakdown of where all of the
compute time was spent (again from the old LAMMPS):
Nbond time/: 6509\.860530 34\.7574
Long time/: 4651.507731 24.8353
Bond time/: 14\.702192 0\.0785
Angle time/: 59.004120 0.3150
Dihed time/: 151\.046199 0\.8065
Impro time/: 2.885573 0.0154
Nay-1 time/: 620\.319198 3\.3120
Nay\-2 time/: 6.750760 0.0360
Exch time/: 20\.119540 0\.1074
Comm time/: 849.082181 4.5334
Fcomm time/: 815\.395753 4\.3536
Shake time/: 814.003810 4.3461
I/O time/: 920\.195881 4\.9131
Other time/: 3294.563788 17.5903
If the communication time (Comm and Fcomm) is
overwhelming, fewer processors may be better, whereas
if Nbond time is large, more processors may be
preferable. Long = long-range electrostatics, so if
this number is large, it may be worth examining ways
to optimize your electrostatics calculation for best
speed at a given level of accuracy (i.e. changing the
real-space cutoff, Ewald vs. PPPM). The bottom line is
that you�ll want to minimize the total simulation time
while maintaining the accuracy.
Paul
--- "Jan-Michael Y. Carrillo"
From: Jan-Michael Y. Carrillo
[mailto:jycarrillo@…19…]Sent: Monday, October 31, 2005 1:58 PM
To: 'Steven J. Plimpton'
Cc: 'Paul Crozier'
Subject: RE: [lammps-users] Documentation for Lammps
OutputThanks Paul and Steve,
My simulation box consists of polyelectrolytes and
counterions and I am
running "coulomb style". I am changing my Coulomb
Cutoff to different
values and running my system using 1, 2 or 3
processors (via MPI)while
maintaining RMS precision/accuracy. I ran my sims
up to 20 MD time
steps only so I can get results immediately. I want
to know which
configuration (cutoff and No. of processors) is best
for my system so
naturally I look in to the screen outputs. I don't
know how to compare
each run ( what I initially did is just compare the
"Total time" output
per run). That is why I am looking for definitions
for the "PPPM timing
info" and other screen outputs pertaining to
processing time, PPPM,
etc...Jan-Michael
From: Steven J. Plimpton [mailto:[email protected]]
Sent: Monday, October 31, 2005 1:12 PM
To: [email protected]...
Cc: [email protected]
Subject: Re: [lammps-users] Documentation for Lammps
Output(Paul can you post this to the list)
>I am using Lammps 2001. By any chance is there some
documentation for
>LAMMPS Output (Output to Screen while the
simulation is running). I can>only see these items in the Lammps Website:
This doc page for the current LAMMPS may be helpful,
since the output is
not that different for 2001 vs 2005.
http://www.cs.sandia.gov/~sjplimp/lammps/doc/Section_start.html#2_5
Thanks a lot Paul,
I did see a trend where "Long Time" decreases as I increase the PPPM
cutoff and noticed that there is lesser time consumed using single
processor. Is this something to do with FFT? Like, is FFT faster using
single processors rather than parallel ones? "Nbond time" increases as I
increase the PPPM Cutoff but decreases as I use more CPU but the
increase/decrease is not as much as the change in Long Time vs. PPPM
Cutoff. Communication time is hardly a factor and doesn't change much.
Its trend though is that as I increase the number of nodes used, I also
increase communication time (Comm + Fcomm). The trend for the "Total
Time" is the same as the trend for the "Long Time". I guess this
indicates that electrostatic interaction computation takes the bulk of
the processing time. And at a certain point, "Total time" for 1, 2 and 3
cpu's converge to a certain cutoff point (indicating that you don't get
any incentive if you use more nodes and single cpu is the optimum).
Finally, I just want to know what are the time units reported in "Total
time" and the rest of the "times" (i.e., Nbond time, long time, loop
time, Fcomm and Comm time)? Are they in seconds?
Jan-Michael
You�re welcome Jan-Michael --- glad I could help. Yes,
what you are observing does have to do with FFTs, but
the observations you�re making for your problem are
not true in general. Sounds like the inter-processor
communication speed on your machine is slow, such that
adding more processors on this problem actually makes
the calculation slower instead of faster. This is not
true for most LAMMPS problems I�ve run using PPPM with
FFTs, although I did see this situation on a set of
polyelectrolyte implicit solvent simulations I did a
few years ago (apparently similar to your problem).
This issue seems particularly problematic for systems
that have highly charged particles with low particle
density. The percentage of time required for the
long-range electrostatics (the �Long time�) is large
for this type of problem. I likewise found, as you are
finding, that using a large real-space cutoff for this
type of problem made the simulation go faster. So what
you are finding coincides with my experience. So it is
a good idea to use a large real-space cutoff on this
type of problem. And there is probably an optimum that
you could find by doing tests like you are doing.
LAMMPS time units are reported in seconds.
Paul