Skip to content

MPI-settings behavior on JSC systems #129

Open

Description

@kvrigor

@AGonzalezNicolas' investigation of eCLM failing on >16 JURECA nodes shows that the different MPI-settings on JSC systems have different memory and performance behavior. Wrong settings can cause failure (as with the linked issue) or constrain model performance. Knowing the optimal MPI-settings can be highly relevant for high-resolution simulations.

The various MPI runtimes are documented here: https://apps.fz-juelich.de/jsc/hps/jureca/software-modules.html#mpi-runtimes

Activity

  1. added
    environmentIssues springing from the environment (in a general sense) where TSMP2 is executed.
    on May 27, 2026
  2. kvrigor commented on Sep 4, 2026

    @kvrigor
    MemberAuthor

    Another instance of MPI runtime error reported by @bnaz2004:

    [jrc0477:45753:0:45753] ib_mlx5_log.c:179  Transport retry count exceeded on mlx5_0:1/IB (synd 0x15 vend 0x81 hw_synd 0/0)
    

    This is caused by too many dropped IP packets across the MPI fabric (1, 2). I'm starting to suspect that very high resolution runs (e.g. DE009, EUR0275) that save NetCDF files in parallel mode (pio_typename = netcdf4p) is pushing the JSC MPI comm fabric to its limit. @AGonzalezNicolas it could be a good idea to benchmark what NetCDF I/O settings are apt given a particular MPI library or runtime setting.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    environmentIssues springing from the environment (in a general sense) where TSMP2 is executed.

    Type

    No type

    Projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions