We study a system that consists of multiple servers, where one has the option to (instantaneously) turn off servers and turn them on (after some delay). Such systems arise in the study of computer server farms where one is interested in simultaneously addressing response times and energy costs. For systems with one queue per server, we show that it may be advantageous for the servers to be heterogeneous in their server control policies. For a system with one queue for the entire system and where the underlying random variables are exponential, we give several structural results for the optimal server control policy. These structural results are then leveraged to yield a tractable means to solve the underlying CTMC. The solution is then used to address a typical design issue: determining the number of servers that should be permanently on. Finally, some open issues and challenges will be discussed.