Dopair vec mpi

This works on EAGLE_25 over 2 nodes. To test the performance can I have the scripts you use to create the MPI scaling plots? Do you do 1 thread per node and scale it up to 16? Or 1 node without MPI and enable MPI after 16 threads?

It's the script you wrote which is located in the examples sub-directory.

I run 1 to 16 threads with the non-MPI version. Then 1 MPI rank per node with 16 threads per rank for the MPI version. Up to 32 nodes.

unset I_MPI_HYDRA_BOOTSTRAP
set I_MPI_EAGER_THRESHOLD=4194304

mpirun -np 32 ../swift_mpi -s -t 16 eagle_25 -a -n 4096

You can also cancel the snapshot writing.

What's the difference between this and !454 (closed) ?

Okay thanks, I will try that.

I created that merge request yesterday by accident. I have deleted it now.

There is a full setup for cosma-6 here: /cosma6/data/dp004/jlvc76/SWIFT/Master/swiftsim/examples/EAGLE_25

unmarked as a Work In Progress

I'll let you check whether just playing with ci_active is enough.

Definitely not harming things. Might be useful for large node counts.

As discussed, this could be simplified by playing with the active status only. Stuff for thoughts after SC...

And so that we don't forget: check whether we can avoid computing indices all together when the other cell is inactive (but local).

added 2 commits

4452cf5e - Only populate index list for cells that are both active and local to the node..
fa153c4c - Comments and formatting.

Compare with previous version

This works for inactive local cells also now

Ok, let me make some changes.

added 1 commit

8e8ed29d - Tidied code up.

Compare with previous version

added 2 commits

7c6f41c5 - Style improvements and use of the faster 'part_is_active()' function consistently.
8e7facf9 - Merge branch 'dopair_vec_MPI' of gitlab.cosma.dur.ac.uk:swift/swiftsim into dopair_vec_MPI

Compare with previous version

resolved all discussions

added 1 commit

61730f1d - More style improvements.

Compare with previous version

Dopair vec mpi

Merge request reports

Activity