In TCPConnectionMap$TCPConnection$Receiver, we create a new byte[] buffer for each received message, based on the length sent before the message.
This buffer is subsequently passed to BasicTCP which calls TP.receive().
TP.receive() makes a copy of the buffer if it is a single message (batches are deserialized in the same thread), before passing the copy to one the thread pools to get deserialized and processed.
This is an unneeded copy.
SOLUTION: add a flag needs_copy to the TP.receive() call and don't copy if true.