To specify using TCP/IP, you need to set I_MPI_DEVICE=ssm. This will run over sockets across nodes, and using the shm device within a node.
Additionally, the error you provide could be due to a failed connection with the node, inability to start the mpd daemon on the remote node, etc. Can you verify that you're using the latest version of Intel MPI Library 3.2 Update 1? You can do so by running "mpiexec -V".
Also, make sure no leftover mpd python processes exist on the nodes. You can do so by running "ps aux | grep mpd". Go ahead and kill any left over mpd.py procs you find.