processors Px Py Pz keyword args ...
Px,Py,Pz = # of processors in each dimension of 3d grid overlaying the simulation domain
zero or more keyword/arg pairs may be appended
keyword = grid or map or part or file
grid arg = gstyle params ... gstyle = onelevel or twolevel or numa or custom onelevel params = none twolevel params = Nc Cx Cy Cz Nc = number of cores per node Cx,Cy,Cz = # of cores in each dimension of 3d sub-grid assigned to each node numa params = none custom params = infile infile = file containing grid layout map arg = cart or cart/reorder or xyz or xzy or yxz or yzx or zxy or zyx cart = use MPI_Cart() methods to map processors to 3d grid with reorder = 0 cart/reorder = use MPI_Cart() methods to map processors to 3d grid with reorder = 1 xyz,xzy,yxz,yzx,zxy,zyx = map processors to 3d grid in IJK ordering numa arg = none part args = Psend Precv cstyle Psend = partition # (1 to Np) which will send its processor layout Precv = partition # (1 to Np) which will recv the processor layout cstyle = multiple multiple = Psend grid will be multiple of Precv grid in each dimension file arg = outfile outfile = name of file to write 3d grid of processors to
processors * * 5 processors 2 4 4 processors * * 8 map xyz processors * * * grid numa processors * * * grid twolevel 4 * * 1 processors 4 8 16 grid custom myfile processors * * * part 1 2 multiple
Specify how processors are mapped as a regular 3d grid to the global simulation box. The mapping involves 2 steps. First if there are P processors it means choosing a factorization P = Px by Py by Pz so that there are Px processors in the x dimension, and similarly for the y and z dimensions. Second, the P processors are mapped to the regular 3d grid. The arguments to this command control each of these 2 steps.
The Px, Py, Pz parameters affect the factorization. Any of the 3 parameters can be specified with an asterisk “*”, which means LAMMPS will choose the number of processors in that dimension of the grid. It will do this based on the size and shape of the global simulation box so as to minimize the surface-to-volume ratio of each processor’s sub-domain.
Choosing explicit values for Px or Py or Pz can be used to override the default manner in which LAMMPS will create the regular 3d grid of processors, if it is known to be sub-optimal for a particular problem. E.g. a problem where the extent of atoms will change dramatically in a particular dimension over the course of the simulation.
The product of Px, Py, Pz must equal P, the total # of processors LAMMPS is running on. For a 2d simulation, Pz must equal 1.
Note that if you run on a prime number of processors P, then a grid such as 1 x P x 1 will be required, which may incur extra communication costs due to the high surface area of each processor’s sub-domain.
Also note that if multiple partitions are being used then P is the number of processors in this partition; see the -partition command-line switch doc page for details. Also note that you can prefix the processors command with the partition command to easily specify different Px,Py,Pz values for different partitions.
You can use the partition command to specify different processor grids for different partitions, e.g.
partition yes 1 processors 4 4 4 partition yes 2 processors 2 3 2
This command only affects the initial regular 3d grid created when the simulation box is first specified via a create_box or read_data or read_restart command. Or if the simulation box is re-created via the replicate command. The same regular grid is initially created, regardless of which comm_style command is in effect.
If load-balancing is never invoked via the balance or fix balance commands, then the initial regular grid will persist for all simulations. If balancing is performed, some of the methods invoked by those commands retain the logical topology of the initial 3d grid, and the mapping of processors to the grid specified by the processors command. However the grid spacings in different dimensions may change, so that processors own sub-domains of different sizes. If the comm_style tiled command is used, methods invoked by the balancing commands may discard the 3d grid of processors and tile the simulation domain with sub-domains of different sizes and shapes which no longer have a logical 3d connectivity. If that occurs, all the information specified by the processors command is ignored.
The grid keyword affects the factorization of P into Px,Py,Pz and it can also affect how the P processor IDs are mapped to the 3d grid of processors.
The onelevel style creates a 3d grid that is compatible with the Px,Py,Pz settings, and which minimizes the surface-to-volume ratio of each processor’s sub-domain, as described above. The mapping of processors to the grid is determined by the map keyword setting.
The twolevel style can be used on machines with multicore nodes to minimize off-node communication. It insures that contiguous sub-sections of the 3d grid are assigned to all the cores of a node. For example if Nc is 4, then 2x2x1 or 2x1x2 or 1x2x2 sub-sections of the 3d grid will correspond to the cores of each node. This affects both the factorization and mapping steps.
The Cx, Cy, Cz settings are similar to the Px, Py, Pz settings, only their product should equal Nc. Any of the 3 parameters can be specified with an asterisk “*”, which means LAMMPS will choose the number of cores in that dimension of the node’s sub-grid. As with Px,Py,Pz, it will do this based on the size and shape of the global simulation box so as to minimize the surface-to-volume ratio of each processor’s sub-domain.
For the twolevel style to work correctly, it assumes the MPI ranks of processors LAMMPS is running on are ordered by core and then by node. E.g. if you are running on 2 quad-core nodes, for a total of 8 processors, then it assumes processors 0,1,2,3 are on node 1, and processors 4,5,6,7 are on node 2. This is the default rank ordering for most MPI implementations, but some MPIs provide options for this ordering, e.g. via environment variable settings.
The numa style operates similar to the twolevel keyword except that it auto-detects which cores are running on which nodes. Currently, it does this in only 2 levels, but it may be extended in the future to account for socket topology and other non-uniform memory access (NUMA) costs. It also uses a different algorithm than the twolevel keyword for doing the two-level factorization of the simulation box into a 3d processor grid to minimize off-node communication, and it does its own MPI-based mapping of nodes and cores to the regular 3d grid. Thus it may produce a different layout of the processors than the twolevel options.
The numa style will give an error if the number of MPI processes is not divisible by the number of cores used per node, or any of the Px or Py of Pz values is greater than 1.
Unlike the twolevel style, the numa style does not require any particular ordering of MPI ranks i norder to work correctly. This is because it auto-detects which processes are running on which nodes.
The custom style uses the file infile to define both the 3d factorization and the mapping of processors to the grid.
The file should have the following format. Any number of initial blank or comment lines (starting with a “#” character) can be present. The first non-blank, non-comment line should have 3 values:
Px Py Py
These must be compatible with the total number of processors and the Px, Py, Pz settings of the processors command.
This line should be immediately followed by P = Px*Py*Pz lines of the form:
ID I J K
where ID is a processor ID (from 0 to P-1) and I,J,K are the processors location in the 3d grid. I must be a number from 1 to Px (inclusive) and similarly for J and K. The P lines can be listed in any order, but no processor ID should appear more than once.
The map keyword affects how the P processor IDs (from 0 to P-1) are mapped to the 3d grid of processors. It is only used by the onelevel and twolevel grid settings.
The cart style uses the family of MPI Cartesian functions to perform the mapping, namely MPI_Cart_create(), MPI_Cart_get(), MPI_Cart_shift(), and MPI_Cart_rank(). It invokes the MPI_Cart_create() function with its reorder flag = 0, so that MPI is not free to reorder the processors.
The cart/reorder style does the same thing as the cart style except it sets the reorder flag to 1, so that MPI can reorder processors if it desires.
The xyz, xzy, yxz, yzx, zxy, and zyx styles are all similar. If the style is IJK, then it maps the P processors to the grid so that the processor ID in the I direction varies fastest, the processor ID in the J direction varies next fastest, and the processor ID in the K direction varies slowest. For example, if you select style xyz and you have a 2x2x2 grid of 8 processors, the assignments of the 8 octants of the simulation domain will be:
proc 0 = lo x, lo y, lo z octant proc 1 = hi x, lo y, lo z octant proc 2 = lo x, hi y, lo z octant proc 3 = hi x, hi y, lo z octant proc 4 = lo x, lo y, hi z octant proc 5 = hi x, lo y, hi z octant proc 6 = lo x, hi y, hi z octant proc 7 = hi x, hi y, hi z octant
Note that, in principle, an MPI implementation on a particular machine should be aware of both the machine’s network topology and the specific subset of processors and nodes that were assigned to your simulation. Thus its MPI_Cart calls can optimize the assignment of MPI processes to the 3d grid to minimize communication costs. In practice, however, few if any MPI implementations actually do this. So it is likely that the cart and cart/reorder styles simply give the same result as one of the IJK styles.
Also note, that for the twolevel grid style, the map setting is used to first map the nodes to the 3d grid, then again to the cores within each node. For the latter step, the cart and cart/reorder styles are not supported, so an xyz style is used in their place.
The part keyword affects the factorization of P into Px,Py,Pz.
It can be useful when running in multi-partition mode, e.g. with the run_style verlet/split command. It specifies a dependency between a sending partition Psend and a receiving partition Precv which is enforced when each is setting up their own mapping of their processors to the simulation box. Each of Psend and Precv must be integers from 1 to Np, where Np is the number of partitions you have defined via the -partition command-line switch.
A “dependency” means that the sending partition will create its regular 3d grid as Px by Py by Pz and after it has done this, it will send the Px,Py,Pz values to the receiving partition. The receiving partition will wait to receive these values before creating its own regular 3d grid and will use the sender’s Px,Py,Pz values as a constraint. The nature of the constraint is determined by the cstyle argument.
For a cstyle of multiple, each dimension of the sender’s processor grid is required to be an integer multiple of the corresponding dimension in the receiver’s processor grid. This is a requirement of the run_style verlet/split command.
For example, assume the sending partition creates a 4x6x10 grid = 240 processor grid. If the receiving partition is running on 80 processors, it could create a 4x2x10 grid, but it will not create a 2x4x10 grid, since in the y-dimension, 6 is not an integer multiple of 4.
If you use the partition command to invoke different “processors” commands on different partitions, and you also use the part keyword, then you must insure that both the sending and receiving partitions invoke the “processors” command that connects the 2 partitions via the part keyword. LAMMPS cannot easily check for this, but your simulation will likely hang in its setup phase if this error has been made.
The file keyword writes the mapping of the factorization of P processors and their mapping to the 3d grid to the specified file outfile. This is useful to check that you assigned physical processors in the manner you desired, which can be tricky to figure out, especially when running on multiple partitions or on, a multicore machine or when the processor ranks were reordered by use of the -reorder command-line switch or due to use of MPI-specific launch options such as a config file.
If you have multiple partitions you should insure that each one writes to a different file, e.g. using a world-style variable for the filename. The file has a self-explanatory header, followed by one-line per processor in this format:
world-ID universe-ID original-ID: I J K: name
The IDs are the processor’s rank in this simulation (the world), the universe (of multiple simulations), and the original MPI communicator used to instantiate LAMMPS, respectively. The world and universe IDs will only be different if you are running on more than one partition; see the -partition command-line switch. The universe and original IDs will only be different if you used the -reorder command-line switch to reorder the processors differently than their rank in the original communicator LAMMPS was instantiated with.
I,J,K are the indices of the processor in the regular 3d grid, each from 1 to Nd, where Nd is the number of processors in that dimension of the grid.
The name is what is returned by a call to MPI_Get_processor_name() and should represent an identifier relevant to the physical processors in your machine. Note that depending on the MPI implementation, multiple cores can have the same name.
This command cannot be used after the simulation box is defined by a read_data or create_box command. It can be used before a restart file is read to change the 3d processor grid from what is specified in the restart file.
The grid numa keyword only currently works with the map cart option.
The part keyword (for the receiving partition) only works with the grid onelevel or grid twolevel options.
The option defaults are Px Py Pz = * * *, grid = onelevel, and map = cart.