3.2. mpi

These parameters directly affect how UMD-LETKF scatters the model state across the processors and how it optimizes file I/O.

Warning

In an attempt to improve memory usage with the MPI calls, I over-optimized, resulting in overly slow performance in the MPI scatter/gather calls if a large domain or high number of ensemble members are used. This will be fixed in the future.

Parameters

ens_size:

type: integer, required

The number of ensemble members.

ppn:

type: integer, default: 1

The number of processors per node.

Optional, but may help improve I/O performance by evenly distributing across nodes the the PEs which perform simultaneous I/O.

Example

mpi:
  ens_size: 20
  ppn: 10