中文摘要：

实时功耗温度管理（DPTM）通过对任务的准确预测与合理调度,可以有效降低片上系统的运行能耗与峰值温度.为了获得更好的DPTM调度效果,文中提出了一种精确的组合式任务预测算法和一种任务调度算法VP-TALK,进而构建了一个完整的DPTM原型系统.为了对复杂任务进行精确的任务预测,文中DPTM系统先将复杂任务按频谱长短分类为随机/周期/趋势3种成分,然后采用灰色模型/傅里叶模型/径向基函数（RBF）神经网络模型分别对这3种成分进行组合分析,以获得精确的预测效果;基于精确预测的任务负载量,文中所提出的VP-TALK算法可以计算出最优电压-频率对的理想值,进而选择出两组与理想值相邻的电压-频率对,以获得两个现实的工作状态,并考虑核心温度和任务实时性的条件,VP-TALK算法将任务负载分配到这两个工作状态,以获得最优的DPTM效果;最后基于机器学习方法,综合4种源算法构建了一套完整的DPTM原型系统.实验结果表明：（1）文中系统的任务预测组合方法的平均误差仅为2.89%;（2）在相同的设定峰值温度约束下,与已有调度算法的能耗值相比,尽管假设了更为敏感的功率-温度影响关系,但对于较高的工作负载率,文中所提出的VP-TALK调度算法仍能够获得平均14.33%的能耗降低;（3）文中所提出的DPTM原型系统可以获得接近于理想状态的能耗优化效果.

英文摘要：

Optimal dynamic power and temperature management（DPTM）methods can effectively cut down the soaring power consumption and alleviate the problem of chip temperature.In order to get better scheduling results,this paper mainly accomplish three things.First,with principles derived from analyzing three previous methods as thumb rules,we obtain an improved DPTM algorithm,named VP-TALK,that carefully schedule the processor＇s running and dormant behaviors.Besides,we propose a combined predicting model.It may predict the workload on the chip so as to draw out optimal but unpractical frequency（F）and voltage（V）.This FV pair decides two distinct pairs of FV,with which VP-TALK schedules the processor according to both the core temperature and remaindering work load.Finally,combining the workload prediction method and four DPTM algorithms,we further build a DPTM control system.Even though ourmodel assume a tighter and more sensitive relationship between energy and temperature,experiments show that（1）the workload prediction＇s error is as less as 2.89%;（2）under even more tough assumptions about thermal and power interrelation and the same peak temperature ceiling value,our proposed DPTM algorithm gains averagely 14.33%energy saving comparing to previous algorithms when the workload ratio is comparatively high;（3）comprehensive DPTM control system＇s managing effect is near to the most ideal one.