When a new node joins large cluster filled with data, it gets the new CH and REBALANCE_START command, and requests data from all nodes at once (or almost all with even distribution of segments). It may be not able to handle this amount of transfers in parallel even at the JGroups level - this results in data sent to the node and discarded at the receiver, sent again and again. With a heavy congestion the node just buffers fragments of a message from one sender and never passes this up.
The number of StateRequestCommands(START_STATE_TRANSFER) should be limited so that the node is not congested.