Results with 1 core does not match with 16 cores for same geometry, model, machine, version, etc.

Unknown
Unknown
edited April 2023 in Community Q&A

Hi experts,

I have a question regarding the outputs we obtain when we use different number of cores.

--> I have a geometry and I ran a simple problem with only 1 core.

--> I copied exactly the same files (along with mesh, input files, etc) to another folder, changed the problem directory location in Acusim.cnf file and changed the number of cores to 16.

Just by doing these things, I see that the value of surface integrated heat flux for the same surface are different for two different runs. Is this likely to happen? Is this happening because of parallelization of the problem by acusolve?

Thanks.

***Interesting***

When I just change the number of cpus in the Acusim.cnf file and run the simulation in the same folder (everything same, nothing changed, just num_processors changed to 1 and then 16), I see different results at different time steps. I have attached the following files.

Tagged:

Answers

  • acupro
    acupro
    Altair Employee
    edited April 2023

    These kinds of differences (less than 1%) would be expected by any solver, not just AcuSolve.

    Do both runs also achieve the same convergence level?

  • Unknown
    Unknown
    edited April 2023
    acupro_21778 said:

    These kinds of differences (less than 1%) would be expected by any solver, not just AcuSolve.

    Do both runs also achieve the same convergence level?

    I think they achieved the same convergence level. 

    Not only this, I rerun the simulation many times by changing convergence tolerance and max times steps in different machines. I tried the same simulation setup in next machine without many any changes (just copied the same files, changed the problem directory in Acusim.cnf), I see some differences in the results.

    The thing is when i use one core in two different machines for same settings, acusolve reports exactly same values at all timesteps (correct to 12 significant digits). But when I just change the num_processors in Acusim.cnf to multicores (in both machines 16), I see difference in results (sometimes it is just correct only upto 1 significant digit). 

    1> Can you specify what could be the sources of these errors?

    2> I would like to achieve at least correct upto 6 significant digits. Is there any way to get this?

    3> Is there any settings to specify the significant digits (may be for dimensions or other parameters)? 

     

    Thanks.

Categories