This requires the following things:

• performance of TCP_NIO2 must be better than TCP
• (possibly) use multiple selectors: each selector handles only a max number of connections; create more selectors when the number of connections per selector exceeds a certain value
• (possibly) get rid of the reader thread and let the above selector do the read (the write is already done by the selector thread)
• Convert Buffers to use off-heap ByteBuffers rather than heap-based ByteBuffers, see JGRP-2241 for details