Model-Forum for Levante

compiler used: Intel OneAPI classic (ifort)

FFLAGS used:

-O3 -fp-model fast=2 -march=core-avx2 -ip –ipo -assume buffered_stdout -r8 -implicitnone

FESOM2 code optimisations:

• MPI optimizations
  • Combine MPI calls for multiple variables using MPI_Type_hindexed_block
  • Persistent communication (MPI_Send_init, MPI_Recv_init)
• Replace MPI_Iallreduce with MPI_Allreduce in PARMS
• Reordered array slope_tapered
• Replace array syntax with loops in update_means
• Precalculated minval(nlevels_nod2d(elem2D_nodes(1:3,el))) in smooth_nod3D

FESOM2 runtime optimisations:

• Solver options can be chosen through added variables in namelist mesh_def
  • pctype=3 # default 2
  • pcilutype=1 # default 1
  • lutype=3 # default 2
  • fillin=2
• Round-robin placement in groups of 8 tasks across NUMA domains.