基于动态规划的暗硅背景下功率受限众核系统芯片功率分配方法研究  

Research on power allocation with dynamic-programming for many-core system in dark silicon era with power-budget-limited

作　　者：何元烈[1] 郑龙[1] 

机构地区：[1]广东工业大学计算机学院,广州510006

出　　处：《计算机应用研究》2016年第4期1079-1082,1086,共5页Application Research of Computers

摘　　要：高性能众核系统芯片快速增长的功耗成为限制其性能发展的瓶颈之一。由于受到散热、封装、外部供电能力等限制,众核系统芯片的功率预算不足以支撑芯片内所有处理器核同时正常工作。针对暗硅背景下功率受限的情况,对众核系统芯片功率分配进行了研究。提出一种新的能量分配算法——基于动态规划的并行分配方式,在功率受限的情况下达到最优化性能。通过实验证明,与其他三种知名的方法相比,该方法可以有效地减少30%的执行时间,同时运行开销和硬件成本也比较小,这使得该方案适用于未来的众核系统。The ever-increasing power consumption has been one of the design challenges for high-performance many-core systems. On-chip processor cores of many-core systems cannot operate at their peak frequency simultaneously due to the limited chip power budget,which is set forth by the chips＇ cooling,packaging,and power supply capacity. This paper formulated the performance optimization problem formally,and proposed power allocation algorithm based dynamic programming to solve this problem,it had a linear time complexity,and it was quite scalable to the problem size. Experimental results have confirmed that this algorithm can reduce the execution time by 30%,compared to other competing methods. The runtime overhead and hardware cost are found to be reasonably small,which makes this scheme suitable for future many-core systems.

关 键 词：并行计算 功率预算 众核系统 动态规划 

分 类 号：TP302.7[自动化与计算机技术—计算机系统结构]