Some nonlinear behavior with MPI
I got some strange behavior when I try to run GPR on the new skylake i9 here in Trento. The application is the GPR one used with the GPR_2D.exahype spec file. The test is a travelling smooth shock and it runs with non static limiter and AMR. I report some results we got:
2D 4 TBB 1MPI 1 AMR NoOpt ok in 177,54s 2D 4 TBB 1MPI 1 AMR Opt ok in 170,85s 2D 8 TBB 1MPI 1 AMR Opt ok in 145,24 2D 2 TBB 1MPI 1 AMR Opt ok in 292,18 2D 2 TBB 4MPI 1 AMR Opt do not work (issue in musclhancock) 2D 1 and 2 TBB 4MPI 1 AMR Opt using Godunov, do not work, it frozen in initialization and doesn't do anything 2D 1 TBB 4MPI 0 AMR NoOpt using Godunov, it explode 2D 1 TBB 1MPI 0 AMR NoOpt ok in 37,42s (every spawn optimization disabled) 2D 1 TBB 1MPI 1 AMR NoOpt ok in 219.433s (every spawn optimization disabled)
In summary it seems to work all the times we use 1 MPI, more than one seems to be in trouble.