In MPI, every send has to have a matching receive, which opens the possibility of receiver-initiated retransmission. In other words, when the receiver is expecting a message and has not received after a certain time, it sends a request for retransmission. While this is in principle possible, we decided for sender-based retransmission. The reason is that MPI allows the use of wild cards, so the receiver might not know from whom to expect the data, and would have to broadcast its retransmission request to all members in the group. They in turn might retransmit packets which have arrived correctly, but had not been acknowledged yet.