While DSP's are now widely adopted in many embedded systems in the cost minimization and the resolving of computing needs of various multimedia applications, little work is done for energy-efficient real-time job scheduling over DSP's. As motivated by the needs, a set of sliding-window-based algorithms are proposed. A sequence of time points and their corresponding processor speeds is generated to run jobs of a periodic task on the DSP, such as that for the decoding of an H.264 stream. An online DVS scheme for energy minimization with constrained buffer size consideration is proposed, and the capability of the scheme is evaluated by a series of experiments over real and synthesized traces. It was shown that the energy saving can be up to 44%, and prediction errors were not significant enough to result in more than 3% in deadline missing.