Runtime Parameters: MPI and OpenMP - hyschive/gamer-fork GitHub Wiki

Parameters described on this page: OMP_NTHREAD,   OPT__INIT_GRID_WITH_OMP,   LB_INPUT__WLI_MAX,   LB_INPUT__PAR_WEIGHT,   OPT__RECORD_LOAD_BALANCE,   OPT__MINIMIZE_MPI_BARRIER  

Parameters below are shown in the format:   Name   (Valid Values)   [Default Value]

  • OMP_NTHREAD   (≥1; <1 → set to default)   [-1]

    • Description: Number of OpenMP threads associated with each MPI process. When enabling MPI, the default is set to the ratio between the total number of CPU cores and the total number of MPI processes. When disabling MPI, the default is set to the maximum number of threads available, which can be controlled by the environment variable OMP_NUM_THREADS. Having OMP_NTHREAD=1 is equivalent to disabling the OpenMP parallelization. See also Hybrid MPI/OpenMP/GPU and MPI Binding and Thread Affinity.
    • Restriction: Only applicable when enabling the compilation option --openmp.

  • OPT__INIT_GRID_WITH_OMP   (0=off, 1=on)   [1]

    • Description: Whether or not to enable OpenMP when assigning the initial condition of different grid patches. In can be enabled in most cases unless, for example, the initial condition setup involves random numbers.
    • Restriction: Only applicable when enabling the compilation option --openmp.

  • LB_INPUT__WLI_MAX   (≥0.0)   [0.1]

    • Description: Weighted load imbalancing (WLI) threshold. Patches on all levels will be redistributed among different MPI processes when the WLI factor is estimated to be higher than a given threshold. See Performance Optimizations: Load Balancing for details.
    • Restriction: Only applicable when enabling the compilation option --mpi.

  • LB_INPUT__PAR_WEIGHT   (≥0.0)   [0.0]

    • Description: Load balancing weight of one particle over one cell. It is used to improve load balancing for the simulations with particles. See Performance Optimizations: Load Balancing for details. The typical values are 1.0 ~ 2.0.
    • Restriction: Only applicable when enabling the compilation options --mpi and --particle.

  • OPT__RECORD_LOAD_BALANCE   (0=off, 1=on)   [1]

    • Description: Record the load balancing information in the log file Record__LoadBalance.
    • Restriction: Only applicable when enabling the compilation option --mpi.

  • OPT__MINIMIZE_MPI_BARRIER   (0=off, 1=on)   [0]

    • Description: Minimize the MPI synchronization between grid and particle routines to improve load balancing. It can improve the performance notably, especially for simulations with particles.
    • Restriction: For simulations with particles, one must enable the compilation option --store_pot_ghost and set PAR_IMPROVE_ACC=1. OPT__TIMING_BALANCE must be disabled. In addition, it is currently recommended to disable AUTO_REDUCE_DT.

Remarks


Links

Add a custom footer
⚠️ **GitHub.com Fallback** ⚠️