When tcp_nodelay is set as part of the broker's address, the performance of the client drops badly, no matter whether the value is set to 'true' or 'false'. I assume that the parameter is being mishandled and is tuning off the tcp_nodelay property (which is by default on) even when set to 'true'.
returns serialised get-put cycles onto a single queue at the rate of 25/sec, but
returns 3017/sec under otherwise identical conditions.
I assume that the lower performance figure is consistent with tcp_nodelay not being active in this test case.