Today the ever-growing energy cost, especially cooling cost of data centers, draws much attention for carbon emission reduction. This paper presents an energy efficient scheduling strategy based on model prediction control (MPC) to reduce cooling cost in data centers. It uses dynamic voltage frequency scaling technology to adjust the frequencies of computing nodes of a cluster in a way to minimize heat recirculation effect among the nodes. The maximum inlet temperature of nodes can be kept under temperature limits with little stable error. The method can also deal with inner disturbance (system model variation) by dynamic frequencies regulation among the nodes. Analysis shows good scalability and small overhead, making the method applicable in huge data centers. A temperature-aware controller is designed to reduce inlet temperatures to improve energy efficiency of data centers. Using a simulated online bookstore run in a heterogeneous data center the proposed method is proved to have larger throughput in both normal and emergency cases compared with existing solutions such as safe least recirculation heat temperature controller and traditional feedback temperature controller. The MPC-based scheduling method also has less inlet temperature and cooling cost comparing with those two methods under same workload.