Skip to content

Exploit batched input in createFromSends

Arndt, Daniel requested to merge github/fork/masterleinad/batched_createFromSends_cleaned into master

Created by: masterleinad

When calling createFromSends in DistributedSearchTreeImpl::communicateResultsBack the input is already batched, but we currently don't try to benefit from that. Given that even for examples that only require very few communication most of the time related to MPI communication is spent in createFromSends (and for communication heavy applications even more so), it might be worth exploiting this. The new overload for sortAndDetermienBufferLayout essentials just computes the correct permutation for the batched ranks and reconstructs the total permutation from that.

I still need to run benchmarks.

Merge request reports