[jdp810]

Hello all,

I am running several jobs using:
parallel_computation.py -n 6 -f solver_settings > solver_out

but when I run more of these (making sure I don't go over the number of cores available) it slows down drastically.

Do you know how to make the cores independent from each other?

Thank you,
JD

[pcg]

What is drastically? Do you get more throughput (simulations per hour lets say) or less?

Assuming that drastically is really drastic, you probably need to look at how MPI binds processes to cores, look into the --bind-to option.

Otherwise, you will never, ever, get that kind of perfect scaling of running two jobs on the same machine in the same time it takes you to run one.
CFD in general is bound by memory bandwidth, not by compute power.
Once you get to 2-3 cores per memory channel you will not be able to go much faster, you may even go slower overall because there will be more pressure on the CPU cache and the CPU will start running at lower frequency.

[jdp810]

Maybe drastically is a little dramatic. The test I've been running (200k cells in 2D, steady, Incompressible RANS, SA w/transition and euler implicit) in 6 cores has an iteration time of around 0.58s. If I start running another one with the same characteristics, the iteration time goes up to 1.35s...

I am using mpich as it is recommended in the SU2 page, compiling the code with:

./meson.py build -Denable-autodiff=true -Dwith-mpi=enabled

and changing the line with mpi_dep in meson to

mpi_dep = [dependency('mpich', required : get_option('with-mpi'))]

How do I use the --bind-to option? Or can you point me to a page where it is explained?

Thanks!

[pcg]

It is a CLI argument of "mpirun", not an SU2 option.
https://www.open-mpi.org/doc/v3.0/man1/mpirun.1.php

I don't know how to pass it to parallel_optimization.py, but that script is not really needed with v7 anyway.

Something else that may explain the slowdown is if you use virtual threads.
That is also generally not good for CFD in general, you should only use physical cores.
There are good discussions about these hardware aspects in the hardware section of CFD-online.