Skip to content

Getting started

The shortest successful gevolution 2.0 run requires:

  • an NVIDIA CUDA-capable GPU;
  • an MPI installation;
  • a GPU-enabled LATfield2 checkout;
  • FFTW3, GSL/CBLAS, and HDF5;
  • a compiled gevolution executable;
  • at least four MPI ranks arranged as -n 2 -m 2.

Minimum process grid

Both process-grid dimensions must be at least two. LATfield2 communication assumes that neighbouring ranks are distinct, so -n 1, -m 1, and single-rank execution are unsupported.

For production runs, assign one MPI rank to each GPU. On a workstation with one GPU, run four ranks on that GPU through NVIDIA Multi-Process Service (MPS).

Choose a first run

First-run checklist

  • nvcc, mpic++, and the required libraries are visible.
  • The selected build configuration uses the correct CUDA_ARCH=sm_XX value.
  • The output directory already exists.
  • The MPI rank count equals n * m.
  • Both n and m are at least two.
  • Ngrid / n is an integer.
  • Ngrid / m is an even integer.
  • Each production rank sees exactly one intended GPU, or MPS is active for a one-GPU workstation run.
  • The startup banner reports version 2.0 alpha, the expected rank count, and the expected OpenMP thread count.