ERROR   49000

Jey
Jey Altair Community Member
edited October 2020 in Community Q&A

 ERROR   49000: Too many parallel processes for the distributed preconditioner.

I reduced ncpus to least 32, (32cores multiples). still I get this error. May I know how to resolve this for MOM solve

Tagged:

Answers

  • Mel
    Mel Altair Community Member
    edited August 2020

    How many unknowns in your model? See near the beginning of the OUT file.

    Is this MoM or MLFMM?

  • Jey
    Jey Altair Community Member
    edited August 2020

    MoM is used.

    image.png.6892a498bc6a2fa30d4234d9f84f2ab7.png

  • Mel
    Mel Altair Community Member
    edited August 2020

    I presume you have not specifically set a preconditioner?

     

    How many processes are you launching in total across all nodes?

  • Mel
    Mel Altair Community Member
    edited August 2020

    The error is usually given for MLFMM when way too many processes for the problem size are used.

    Have tried just maybe using 32 or 64 processes or so?

    Could you attach the full out file with error?

  • Jey
    Jey Altair Community Member
    edited August 2020

    Not sure on preconditioner settings used.

    Number of parallel processes:     128 on 8 CPUs

    <?xml version="1.0" encoding="UTF-8"?>image.thumb.png.24ec21880673fe8f01e2c61495e52122.png

  • Mel
    Mel Altair Community Member
    edited August 2020

    Ok, this is a hybrid FEM/MoM solution, I missed the FEM tetrahedra in your 2nd post...

     

    Because the FEM part is so small, you will have to reduce the number of processes.

    Try 64, else try 32 processes.

     

    Else if the FEM part does not contain too much fine details (very fine geometry or multiple dielectrics) you could change the FEM part to use surface meshing.
    Then it will be a MoM only solution.

  • Jey
    Jey Altair Community Member
    edited August 2020
  • Mel
    Mel Altair Community Member
    edited August 2020

    Actually, you could enable the MLFMM also and try FEM/MLFMM as the MoM part is quite large.

     

  • Jey
    Jey Altair Community Member
    edited August 2020

    Performing comparison study on FEM-IE  TO FEM-MoM. MoM is structurally large compared to FEM part in the problem.

    MLFMM will be second choice,if MoM could not succeed.

    MoM - run with 32/64 core has out-of-core, memory issue.

    Any suggestions ?

  • Mel
    Mel Altair Community Member
    edited August 2020

    If you have multiple nodes, use only say 8 processes per node, so maybe 64 processes spread out over 8 nodes.

     

    If there is symmetry in the model, apply electric or magnetic symmetry. The MoM memory will reduce by a factor 2 or 4 per symmetry plane.

     

    You can also try to increase the size of the FEM part (to allow using more processes) by meshing the FEM finer (I see there are warnings about the tetrahedra sizes)

    Or you could add air dielectric around the FEM part which would also reduce the size of the FEM/MoM boundary and thus the memory. See videos here and here.