The RMI transport accepts connections in one thread and then spawns a new
thread to accept the next connection. This accept strategy sometimes
ends up refusing connections to clients if client requests happen in a burst
that is larger than the TCP queue (currently the ServerSocket defaults to
a backlog of 50 incoming connections).

Several things can be done to mitigate connection refusal under bursty
conditions. One thing to do is increase the length of the queue when a
ServerSocket is created (100 perhaps). The other thing to do is to
accept in a single thread and spawn a thread to handle the connection
rather than spawning a thread to do the accept. This will result in a
slower handling messages received on a connection since the connection
establishment must wait until the thread servicing the connection is
started. However, handling the accepting in a thread will allow much
faster servicing of connection requests thus preventing clients from
experiencing "connection refused" in many circumstances.