Looks like there is a starvation issue when using "wild" permanent recv requests, i.e. recv requests created with MPI_Recv_init(...,...,..,MPI_ANY_SOURCE, ... It appears to me that this requests are more likely matched with send operations of low ranked peers, resulting in a possible starvation for higher ranked peers, especially in large jobs. At least, this is what I also observed with ITCA.
I observed this with intel-mpi/64/4.0.3/008
Could some Intel developer check or comment on this?