Search

link to homepage

Institute for Advanced Simulation (IAS)

Navigation and service


Why is MPI_Allgatherv (using non-contiguous memory) so slow?

Using non-contiguous memory for MPI_Allgatherv potentially results in
significant performance loss. The easiest way to fix this is the following

export PAMID_COLLECTIVE_ALLGATHERV=GLUE_BCAST

runjob --exp-env PAMID_COLLECTIVE_ALLGATHERV ...

This handles non-contiguous memory for MPI_Allgatherv faster than the
algorithm which is used by default. For larger scales the performance is
still slower than for contiguous memory so you might want to avoid using
non-contiguous memory.


Servicemeu

Homepage