Open jeffhammond opened 5 years ago
AVL tree will be faster than a linked list when the number of memory regions becomes large enough. This optimization was considered and not implemented in the early days of ARMCI-MPI because we didn't see cases where there would be enough performance benefit to justify the increase in complexity. The situation may be different now, but I wanted to make sure this rationale is captured so the current implementation is not considered to be a performance bug.
Yes, I plan to support AVL as an option, in addition to forwards and backwards linear traversal. We can see which one wins for different workloads.
Slab suballocation will eliminate lookup overhead, which is also on my TODO list.
AVL tree will be faster than linked-list traversal for GMR lookup.
@jdinan noted:
Obsolete comment by @jeffhammond:
Migrated from https://github.com/jeffhammond/armci-mpi/issues/25