More embedded system designers pay attention to how to reduce the po wer consumption of I/O devices. Traditional dynamic power management (DPM) polic ies only focus on the device requests, and neglect the application features behi nd the workload. Because of the assumption about the stationary workload, tradit ional DPM policies can not reach their expected goal under the multi-task envir o nment. The paper presents a stack-based predictive timeout strategy (SBPT). It c an predict the access pattern of the device I/O operations by analyzing the ca lling and stack information of tasks and combine predictions of multiple tasks to form the global prediction according to the multiple-service-requester mode l. At last, classify the I/O request by the global prediction and then make the de cision with the timeout technique based on the distribution of the grouped reque sts. An evaluation study of SBPT using the trace-driven simulation is performed. The results show that SBPT can adapt the non-stationary multi-task environment and reduces power consumption more efficiently than other policies.