run_style style args
style = verlet or verlet/split or respa or respa/omp
verlet args = none verlet/split args = none respa args = N n1 n2 ... keyword values ... N = # of levels of rRESPA n1, n2, ... = loop factors between rRESPA levels (N-1 values) zero or more keyword/value pairings may be appended to the loop factors keyword = bond or angle or dihedral or improper or pair or inner or middle or outer or hybrid or kspace bond value = M M = which level (1-N) to compute bond forces in angle value = M M = which level (1-N) to compute angle forces in dihedral value = M M = which level (1-N) to compute dihedral forces in improper value = M M = which level (1-N) to compute improper forces in pair value = M M = which level (1-N) to compute pair forces in inner values = M cut1 cut2 M = which level (1-N) to compute pair inner forces in cut1 = inner cutoff between pair inner and pair middle or outer (distance units) cut2 = outer cutoff between pair inner and pair middle or outer (distance units) middle values = M cut1 cut2 M = which level (1-N) to compute pair middle forces in cut1 = inner cutoff between pair middle and pair outer (distance units) cut2 = outer cutoff between pair middle and pair outer (distance units) outer value = M M = which level (1-N) to compute pair outer forces in hybrid values = M1 M2 ... (as many values as there are hybrid sub-styles M1 = which level (1-N) to compute the first pair_style hybrid sub-style in M2 = which level (1-N) to compute the second pair_style hybrid sub-style in M3,etc kspace value = M M = which level (1-N) to compute kspace forces in
run_style respa 4 2 2 2 bond 1 dihedral 2 pair 3 kspace 4
run_style respa 4 2 2 2 bond 1 dihedral 2 inner 3 5.0 6.0 outer 4 kspace 4
run_style respa 3 4 2 bond 1 hybrid 2 2 1 kspace 3
Choose the style of time integrator used for molecular dynamics simulations performed by LAMMPS.
The verlet style is the velocity form of the Stoermer-Verlet time integration algorithm (velocity-Verlet)
The verlet/split style is also a velocity-Verlet integrator, but it splits the force calculation within each timestep over 2 partitions of processors. See the -partition command-line switch for info on how to run LAMMPS with multiple partitions.
Specifically, this style performs all computation except the kspace_style portion of the force field on the first partition. This include the pair style, bond style, neighbor list building, fixes including time integration, and output. The kspace_style portion of the calculation is performed on the second partition.
This can lead to a significant speedup, if the number of processors can be easily increased and the fraction of time is spent in computing Kspace interactions is significant, too. The two partitions may have a different number of processors. This is most useful for the PPPM kspace_style when its performance on a large number of processors degrades due to the cost of communication in its 3d FFTs. In this scenario, splitting your P total processors into 2 subsets of processors, P1 in the first partition and P2 in the second partition, can enable your simulation to run faster. This is because the long-range forces in PPPM can be calculated at the same time as pairwise and bonded forces are being calculated and the parallel 3d FFTs can be faster to compute when running on fewer processors. Please note that the scenario of using fewer MPI processes to reduce communication overhead can also be implemented through using MPI with OpenMP threads via the INTEL, KOKKOS, or OPENMP package. This alternative option is typically more effective in case of a fixed number of available processors and less complex to execute.
To use the verlet/split style, you must define 2 partitions with the
-partition command-line switch, where partition P1
is either the same size or an integer multiple of the size of the
partition P2. Typically having P1 be 3x larger than P2 is a good
choice, since the (serial) performance of LAMMPS is often best if the
time spent in the
Pair computation versus
Kspace is a 3:1 split.
The 3d processor layouts in each partition must overlay in the following
sense. If P1 is a Px1 by Py1 by Pz1 grid, and P2 = Px2 by Py2 by Pz2,
then Px1 must be an integer multiple of Px2, and similarly for Py1 a
multiple of Py2, and Pz1 a multiple of Pz2.
Typically the best way to do this is to let the first partition choose its own optimal layout, then require the second partition’s layout to match the integer multiple constraint. See the processors command with its part keyword for a way to control this, e.g.
processors * * * part 1 2 multiple
You can also use the partition command to explicitly specify the processor layout on each partition. E.g. for 2 partitions of 60 and 15 processors each:
partition yes 1 processors 3 4 5
partition yes 2 processors 3 1 5
When you run in 2-partition mode with the verlet/split style, the thermodynamic data for the entire simulation will be output to the log and screen file of the first partition, which are log.lammps.0 and screen.0 by default; see the -plog and -pscreen command-line switches to change this. The log and screen file for the second partition will not contain thermodynamic output beyond the first timestep of the run.
See the Accelerator packages page for performance details of the speed-up offered by the verlet/split style. One important performance consideration is the assignment of logical processors in the 2 partitions to the physical cores of a parallel machine. The processors command has options to support this, and strategies are discussed in Section 5 of the manual.
The respa style implements the rRESPA multi-timescale integrator (Tuckerman) with N hierarchical levels, where level 1 is the innermost loop (shortest timestep) and level N is the outermost loop (largest timestep). The loop factor arguments specify what the looping factor is between levels. N1 specifies the number of iterations of level 1 for a single iteration of level 2, N2 is the iterations of level 2 per iteration of level 3, etc. N-1 looping parameters must be specified.
Thus with a 4-level respa setting of “2 2 2” for the 3 loop factors, you could choose to have bond interactions computed 8x per large timestep, angle interactions computed 4x, pair interactions computed 2x, and long-range interactions once per large timestep.
The timestep command sets the large timestep for the outermost rRESPA level. Thus if the 3 loop factors are “2 2 2” for 4-level rRESPA, and the outer timestep is set to 4.0 fs, then the inner timestep would be 8x smaller or 0.5 fs. All other LAMMPS commands that specify number of timesteps (e.g. thermo for thermo output every N steps, neigh_modify delay/every parameters, dump every N steps, etc) refer to the outermost timesteps.
The rRESPA keywords enable you to specify at what level of the hierarchy various forces will be computed. If not specified, the defaults are that bond forces are computed at level 1 (innermost loop), angle forces are computed where bond forces are, dihedral forces are computed where angle forces are, improper forces are computed where dihedral forces are, pair forces are computed at the outermost level, and kspace forces are computed where pair forces are. The inner, middle, outer forces have no defaults.
For fixes that support it, the rRESPA level at which a given fix is active, can be selected through the fix_modify command.
The inner and middle keywords take additional arguments for cutoffs that are used by the pairwise force computations. If the 2 cutoffs for inner are 5.0 and 6.0, this means that all pairs up to 6.0 apart are computed by the inner force. Those between 5.0 and 6.0 have their force go ramped to 0.0 so the overlap with the next regime (middle or outer) is smooth. The next regime (middle or outer) will compute forces for all pairs from 5.0 outward, with those from 5.0 to 6.0 having their value ramped in an inverse manner.
Note that you can use inner and outer without using middle to split the pairwise computations into two portions instead of three. Unless you are using a very long pairwise cutoff, a 2-way split is often faster than a 3-way split, since it avoids too much duplicate computation of pairwise interactions near the intermediate cutoffs.
Also note that only a few pair potentials support the use of the inner and middle and outer keywords. If not, only the pair keyword can be used with that pair style, meaning all pairwise forces are computed at the same rRESPA level. See the doc pages for individual pair styles for details.
Another option for using pair potentials with rRESPA is with the hybrid keyword, which requires the use of the pair_style hybrid or hybrid/overlay command. In this scenario, different sub-styles of the hybrid pair style are evaluated at different rRESPA levels. This can be useful, for example, to set different timesteps for hybrid coarse-grained/all-atom models. The hybrid keyword requires as many level assignments as there are hybrid sub-styles, which assigns each sub-style to a rRESPA level, following their order of definition in the pair_style command. Since the hybrid keyword operates on pair style computations, it is mutually exclusive with either the pair or the inner/middle/outer keywords.
When using rRESPA (or for any MD simulation) care must be taken to choose a timestep size(s) that ensures the Hamiltonian for the chosen ensemble is conserved. For the constant NVE ensemble, total energy must be conserved. Unfortunately, it is difficult to know a priori how well energy will be conserved, and a fairly long test simulation (~10 ps) is usually necessary in order to verify that no long-term drift in energy occurs with the trial set of parameters.
With that caveat, a few rules-of-thumb may be useful in selecting respa settings. The following applies mostly to biomolecular simulations using the CHARMM or a similar all-atom force field, but the concepts are adaptable to other problems. Without SHAKE, bonds involving hydrogen atoms exhibit high-frequency vibrations and require a timestep on the order of 0.5 fs in order to conserve energy. The relatively inexpensive force computations for the bonds, angles, impropers, and dihedrals can be computed on this innermost 0.5 fs step. The outermost timestep cannot be greater than 4.0 fs without risking energy drift. Smooth switching of forces between the levels of the rRESPA hierarchy is also necessary to avoid drift, and a 1-2 Angstrom “healing distance” (the distance between the outer and inner cutoffs) works reasonably well. We thus recommend the following settings for use of the respa style without SHAKE in biomolecular simulations:
run_style respa 4 2 2 2 inner 2 4.5 6.0 middle 3 8.0 10.0 outer 4
With these settings, users can expect good energy conservation and roughly a 2.5 fold speedup over the verlet style with a 0.5 fs timestep.
If SHAKE is used with the respa style, time reversibility is lost, but substantially longer time steps can be achieved. For biomolecular simulations using the CHARMM or similar all-atom force field, bonds involving hydrogen atoms exhibit high frequency vibrations and require a time step on the order of 0.5 fs in order to conserve energy. These high frequency modes also limit the outer time step sizes since the modes are coupled. It is therefore desirable to use SHAKE with respa in order to freeze out these high frequency motions and increase the size of the time steps in the respa hierarchy. The following settings can be used for biomolecular simulations with SHAKE and rRESPA:
fix 2 all shake 0.000001 500 0 m 1.0 a 1
run_style respa 2 2 inner 1 4.0 5.0 outer 2
With these settings, users can expect good energy conservation and roughly a 1.5 fold speedup over the verlet style with SHAKE and a 2.0 fs timestep.
For non-biomolecular simulations, the respa style can be advantageous if there is a clear separation of time scales - fast and slow modes in the simulation. For example, a system of slowly-moving charged polymer chains could be setup as follows:
run_style respa 2 8
This is two-level rRESPA with an 8x difference between the short and long timesteps. The bonds, angles, dihedrals will be computed every 0.5 fs (assuming real units), while the pair and kspace interactions will be computed once every 4 fs. These are the default settings for each kind of interaction, so no additional keywords are necessary.
Even a LJ system can benefit from rRESPA if the interactions are divided by the inner, middle and outer keywords. A 2-fold or more speedup can be obtained while maintaining good energy conservation. In real units, for a pure LJ fluid at liquid density, with a sigma of 3.0 Angstroms, and epsilon of 0.1 kcal/mol, the following settings seem to work well:
run_style respa 3 3 4 inner 1 3.0 4.0 middle 2 6.0 7.0 outer 3
The respa/omp style is a variant of respa adapted for use with pair, bond, angle, dihedral, improper, or kspace styles with an omp suffix. It is functionally equivalent to respa but performs additional operations required for managing omp styles. For more on omp styles see the Speed omp doc page. Accelerated styles take the same arguments and should produce the same results, except for round-off and precision issues.
See the Accelerator packages page for more instructions on how to use the accelerated styles effectively.
The verlet/split style can only be used if LAMMPS was built with the REPLICA package. Correspondingly the respa/omp style is available only if the OPENMP package was included. See the Build package page for more info. It is not compatible with kspace styles from the INTEL package.
Whenever using rRESPA, the user should experiment with trade-offs in speed and accuracy for their system, and verify that they are conserving energy to adequate precision.
For run_style respa, the default assignment of interactions to rRESPA levels is as follows:
bond forces = level 1 (innermost loop)
angle forces = same level as bond forces
dihedral forces = same level as angle forces
improper forces = same level as dihedral forces
pair forces = level N (outermost level)
kspace forces = same level as pair forces
inner, middle, outer forces = no default
(Tuckerman) Tuckerman, Berne and Martyna, J Chem Phys, 97, p 1990 (1992).