FANS install via conda

[sanathkeshav]

Although FANS can be installed via conda super easily -> to be run in parallel via mpi - one needs to install a mpi dist to obtain mpiexec.

I know it's very trivial, but the documentation should also mention this, in my opinion.

[sanathkeshav]

I have also observed some warnings when I use the conda binary:

No protocol specified
[daework3:2960149] shmem: mmap: an error occurred while determining whether or not /tmp/ompi.daework3.2006/jf.0/2999123968/shared_mem_cuda_pool.daework3 could be created.
[daework3:2960149] create_and_attach: unable to create shared memory BTL coordinating structure :: size 134217728 
--------------------------------------------------------------------------
A system call failed during sm BTL initialization that should
not have.  It is likely that your MPI job will now either abort or
experience performance degradation.

  System call: open(2)
  Error:       No such file or directory (errno 2)
--------------------------------------------------------------------------
--------------------------------------------------------------------------

But FANS runs normally after the warning messages (perhaps with a performance degradation).
Do not have a fix at the moment.. any ideas?

[claudiushaag]

The conda-package actually installs openmpi. So doing something like

pixi init
pixi add fans
pixi s
which mpiexec

in an arbitrary empty folder reveals the installed mpiexec command.
When exposing just the binary, like with pixi global, the situation is different. One could additionally expose the mpiexec command from that as a quick fix.
Alternatively, I would wonder why FANS does not call mpiexec by itself as a CLI-call like fans parallel or something l ike that...?

I'm currently looking into the problem creating the shared memory BTL, this also seems to be a problem on my end. That might be connected with the recipe's openmpi. @sanathkeshav, Did your problem occur when using your native openmpi-executable?