[OMPI devel] Connection architectures behind the MPI

[张晶]

In the MPI_THREAD_MULTIPLE mode ,every thread can use MPI_Send(),
MPI_Recv() simulately. I wonder the connection architectures behind the
implement。Does every thread have its own specific socket storage？Or all the
thread share the same socket when they connect to the same rank process and
use some synchronization  mechanism。Is there any paper relate to this topic?