How to find the time taken for forcefield calculations?

LAMMPS LAMMPS Mailing List Mirror
1

Hello all,

I am using two forcefields in my simulation. One is REBO and another Lennard-Jones. I want to compare how much time LAMMPS takes to calculate each of these forcefields in my simulation. Is there any way to reach my goal?

Thanks,

Farshad

2

Hello all,

I am using two forcefields in my simulation. One is REBO and another
Lennard-Jones. I want to compare how much time LAMMPS takes to calculate
each of these forcefields in my simulation. Is there any way to reach my
goal?

*Always* read the documentation before asking such a trivial question like this.

http://lammps.sandia.gov/doc/Section_start.html#start_8

axel.

3

Thanks so much Axel for you response.

I looked at my log file and I saw that the CPU time percentage of all pair potentials (REBO+Lennard-Jones) are summed together as 57% and Kspace as 36% at the end of the file. Although it is helpful, I want to know how much of the 57% is taken for REBO and how much for LJ. To do so, it seems that I should modify verlet.cpp, timer.cpp and finish.cpp files.

I think I should start with verlet.cpp file and edit the following lines:

if (pair_compute_flag) {
force->pair->compute(eflag,vflag);
timer->stamp(TIME_PAIR);
}

But, I don’t how to separately call the timer->stamp function for REBO and LJ. Can anyone help me?

4

Thanks so much Axel for you response.

I looked at my log file and I saw that the CPU time percentage of all pair potentials (REBO+Lennard-Jones) are summed together as 57% and Kspace as 36% at the end of the file. Although it is helpful, I want to know how much of the 57% is taken for REBO and how much for LJ. To do so, it seems that I should modify verlet.cpp, timer.cpp and finish.cpp files.

No.

​ That won’t do and even if it would work, it may not provide any more useful information beyond what you already have.​
please read on…
​​

I think I should start with verlet.cpp file and edit the following lines:

if (pair_compute_flag) {
force->pair->compute(eflag,vflag);
timer->stamp(TIME_PAIR);
}

But, I don’t how to separately call the timer->stamp function for REBO and LJ. Can anyone help me?

You obviously run with a hybrid pair_style, thus you would have to modify that

​style itself ​
and do your own accumulators.

But even then it is not so simple. If you run in parallel, each MPI rank may have a different distribution
​between the various sub styles ​
depending on the​chosen ​
domain decomposition​ and the structure of your system​
. Plus you may also have​to account for ​
load imbalances​ (which may be spread out over multiple force computations)​
. So breaking down the timing would not be too helpful unless you collect data for each MPI rank individually​ and quantify the wait times for any synchronizations​
.​It​
may be simpler to use a profiling tool like perf​ (which works on a per process basis and is Linux-only, but doesn’t require instrumentation)​ or use a tool TAU (which requires a lot of experience to be used well and thus is not recommended for the casual user).

​axel.​

5

You could get a simple estimate, by using
the rerun command and two scripts, each
of which invokes only one of the two pair styles.

Re-process a few snapshots and see how much

time is spent in the pair style by each script.

Steve