Enable MPI by default?

[willend]

Most of us are almost running parallel simulations "by default" - realities of modern computing hardware...

This means that the default behaviour of mcrun to not enable MPI feels somewhat old-fashioned? Let alone the annoyance of having built a binary using the default, subsequently running an mpi-enabled run without MPI compiled in...

Should we change the default to always compile MPI support in?

[mads-bertelsen]

That would be great, I think the work you have done on distribution and install of the package has made it very rare that MPI doesn't work, so the downsides are minimal. Would you still have the default number of cores be 1 but just always compile it with mpi support?

[willend]

Yes, this would be the point. Running "with mpi" and a single process in practice is ~ "not running with mpi"

[mads-bertelsen]

A default value of 2 could also be considered, but probably best to first make the transition with just compiling

[willend]

The PR includes a taste of both:

• mcrun enables compilation with MPI by default, user has to explicitly set to --auto or some number to execute in MPI mode
• mcgui is now set up for compiling MPI in (as per. mcrun) and already has 'auto' as the default number of threads -> so this runs on the full machine

The most serious "consequence" we will get from this possible new default is that 'auto' in Monitor_nD will more or less consequently break... But this could be worked around by saving a particle list buffer and only apply the auto-binning/limits at save-time

[willend]

As easy as this... https://github.com/mccode-dev/McCode/tree/mpi-by-default

@mads-bertelsen @tkittel @g5t @ebknudsen @t-weber @farhi @jorobledo good or bad idea? :-)

[tkittel]

Good idea, as long as it does not actually consume more than 1 thread by default (perhaps I could be convinced of 2).

[willend]

@tkittel agreed. My thinking was also simply to bake in the possibility to not require recompilation for access to multi-thread execution.

[g5t]

mccode-antlr does not control the environment that it runs in and I think its more likely for a system to have a C compiler than MPI, plus I've recently introduced a different way to avoid running the wrong binary, so mcrun-antlr will likely not introduce the same default.

This will be a divergence between McCode and mccode-antlr; but I don't think that is a huge issue since it is only the default behavior.