IMD

Usage

After having compiled IMD with the suitable options, the simulation is started as follows:

    imd_program [-r N] -p parameter_file

The name of the IMD program is the same as the compilation target. If the parameter -r N is present, where N is an integer number, the simulation is restarted from the checkpoint number N. The parameter file contains all further parameters describing the simulation.

Parallel execution

IMD supports two parallelisation methods, OpenMP and MPI, which can also be combined. If OpenMP is used, the number of threads in each process has to specified by setting the envirent variable OMP_NUM_TREADS. If MPI is used, the number of MPI processes must be stated at the program start. This is usually done with a (system dependent) startup script like the following ones:

    mpirun -np P imd_program [-r N] -p parameter_file   # MPICH
    mpprun -n  P imd_program [-r N] -p parameter_file   # Cray T3E

Here, P is the number of MPI processes. If P does not correspond to the number of processes specified by the parameter cpu_dim in the parameter file or if cpu_dim is missing, IMD chooses an appropriate cpu grid by factorizing P evenly and by adhering to the size distribution of cpu_dim (if given).

If MPI and OpenMP are combined, one usually wants to run one MPI process per computer node, with the number of OpenMP threads equal to the number of CPUs per node. Care must be taken that exactly one process is allocated on each node. In a pure MPI scheme, the number of processes P should be equal to the total number of CPUs available. Which parallelisation method is faster has to be checked case by case.