m-a-d-n-e-s-s / madness

Multiresolution Adaptive Numerical Environment for Scientific Simulation
GNU General Public License v2.0
181 stars 62 forks source link

introduced `WorldGopInterface::{set_,}max_reducebcast_msg_size()` #520

Closed evaleev closed 10 months ago

evaleev commented 10 months ago

They control the max size of buffers used by reduce/broadcast ... this may help the issues with MPI errors seen on Perlmutter when sum-reducing large arrays.

Control programmatically (world.gop.set_max_reducebcast_msg_size()) as needed or once at startup by setting environment variable MAD_MAX_REDUCEBCAST_MSG_SIZE

evaleev commented 10 months ago

might be related to the issue this PR tries to work around: https://github.com/open-mpi/ompi/issues/6091