Information theory has shown that feedback can greatly improve the error probability performance of communication channels. In practical communication systems with random message arrivals and queuing, however, this error improvement may come at the expense of longer delay and additional burstiness caused by the retransmissions. Motivated by this, we assess the benefits of using feedback from an integrated information-theoretic and queuing-theoretic perspective. In particular, we characterize policies which achieve the optimal tradeoff between transmission power and packet queuing delay for communication over fading channels with feedback. It is shown that the use of feedback yields substantial improvements in the power-delay tradeoff over systems without feedback.