Memory-lean implementation of FOF
New strategy:
- Run FOF only on the linkables.
- Do all the union-finding business also over MPI.
- Assign the group IDs to all the linkables.
- (new) Run a new FOF loop over the particles this time on the attachable and find their nearest linkable. Record the distance for this and assign the linkable's groupID to the attachable.
- At this stage each particle knows which group it is in.
- Allocate arrays on each node to collect all the group info by looping over every particle.
- All-reduce these arrays.
- If needed, seed black holes.
- As all nodes now have all the info about every group, simplify the i/o and have only rank 0 dump the file.
Edited by Matthieu Schaller
Merge request reports
Activity
added performance label
assigned to @matthieu
added 75 commits
-
3cecaa89...e5cb1abe - 73 commits from branch
master
- 83b3c44a - Merge branch 'master' into leaner_fof
- 6fd146e8 - Remove unwanted functions
-
3cecaa89...e5cb1abe - 73 commits from branch
added 1 commit
- 83dbb2e2 - Version that can compute the correct group size and group mass. Writing...
added 1 commit
- c19deb14 - First version of the distributed CoM computation
added 1 commit
- b7578287 - Compute the local and global number of seeds
added 1 commit
- f1055cbd - Remove unnecessary components of the hashmap
Please register or sign in to reply