Closed
Description
Per #6976, the PSM2 transport in the OFI MTL does not handle messages larger than 4GB bytes (other OFI providers do). The verbs libfabric provider has a similar issue -- it is limited to 2GB messages.
PR's on master/v4.0.x/v3.1.x/v3.0.x were added to make the ofi MTL obey the libfabric property specifying the max message size (ep_attr->max_msg_size
): #7003, #7004, #7005, #7006. This at least makes Open MPI noisly fail to send large messages in this case (vs. silently fail, which is what it was doing before).
This is a good stop-gap solution, but the proper solution is to allow MPI applications to send "arbitrarily-sized" messages (i.e., messages only limited by resources such as memory space).