面向高性能计算机系统的动态电压调节技术研究与能耗优化算法实现

发布时间：2018-06-16 10:16

本文选题：动态电压调节 + 低功耗　；参考：《国防科学技术大学》2012年硕士论文

【摘要】：提高计算系统效能是当前高性能计算领域最为关注的问题之一。系统的高功耗导致巨额的系统运转费用，增加了芯片制造成本，需要更高效的散热技术，并且严重影响系统的可靠性和稳定性。因此，高性能计算机系统需要研究低功耗优化技术，解决日益严峻的功耗问题。动态电压调节（DVS）技术是目前微处理器普遍支持的一种低功耗技术，它以延长执行时间为代价来达到减少处理器等部件能耗的目的，但时间延长将导致系统其它部件的能耗上升，程序运行的能耗的并不一定减小。对于高性能计算系统，由于其规模庞大、维护复杂，在计算任务完成后系统节点往往不会关闭电源或休眠，而是转入空转状态，该状态下的系统的功率仍然很高。为对比DVS技术节能效果，针对高性能计算机系统存在空转时间的运行特点，论文提出了考虑空转能耗影响的系统能耗模型（SEM），，该模型不只考虑程序运行时的能耗，而是考虑相同时间内系统整体的能量消耗，其中包括了空转时系统的能耗。在不同应用系统上进行能耗测试结果显示，采用SEM更有利于对DVS技术的节能效果的评估。基于SEM，论文提出了时间约束条件下的能耗优化模型（EOM），该模型将程序看作由一系列区域组合而成，假定已知每个区域在不同频率下的运行时间和能耗，在满足运行时间限定的前提下，找到一个合适区域并将其以适当频率运行来优化系统的整体能耗。对于把整个程序作为EOM模型中的唯一区域考虑的情况，论文将EOM模型进行简化得到简单能耗优化模型（SEOM），并且在不同应用系统上实现了该模型。实验测试结果显示，对于并行程序，SEOM最大能够获得16%的能耗节省。对于把程序的函数结构和外层循环结构作为EOM候选区域的情况，论文将EOM模型近似转化为复杂能耗优化模型（CEOM）。CEOM通过插桩的手段获取程序各区域的运行信息和植入DVS代码，优化程序运行能耗。基于GCC编译器实现了串行程序的CEOM，实验结果显示根据不同的时间约束条件和程序，CEOM可以获得最大12%的系统能耗节省。
[Abstract]:Improving the efficiency of computing system is one of the most concerned problems in the field of high performance computing. The high power consumption of the system leads to the huge operating cost of the system, increases the cost of chip manufacturing, requires more efficient heat dissipation technology, and seriously affects the reliability and stability of the system. Therefore, high performance computer systems need to study low power optimization technology to solve the increasingly serious power problem. Dynamic Voltage regulator (DVS) is a kind of low power technology which is widely supported by microprocessors at present. It can reduce the energy consumption of processors and other components at the cost of extended execution time, but the time extension will lead to the increase of energy consumption of other components of the system. The energy consumption of the program does not necessarily decrease. For high performance computing system, because of its large scale and complex maintenance, the system nodes often do not turn off power supply or sleep after the calculation task is completed, but switch to idling state, and the power of system in this state is still very high. In order to compare the energy saving effect of DVS technology, aiming at the running characteristic of idle time in high performance computer system, a system energy consumption model considering the influence of idle energy consumption is put forward in this paper, which not only considers the energy consumption of program running. The overall energy consumption of the system at the same time is considered, including the energy consumption of the system when it is idling. The results of energy consumption test on different application systems show that SEM is more helpful to evaluate the energy saving effect of DVS technology. Based on SEM, an energy consumption optimization model with time constraints is proposed. The model considers the program as a series of regions, assuming that the operating time and energy consumption of each region is known at different frequencies. In order to optimize the overall energy consumption of the system, a suitable region is found and run at a suitable frequency under the premise of satisfying the operating time limit. For the case that the whole program is considered as the only region in the EOM model, the EOM model is simplified to obtain a simple energy consumption optimization model, and the model is implemented on different application systems. The experimental results show that for parallel program SEOM, the maximum energy saving is 16%. When the function structure and outer loop structure of the program are regarded as candidate areas of EOM, the EOM model is approximately transformed into a complex energy consumption optimization model, CEOMU. CEOM can obtain the operation information of each area of the program and implant DVS code by inserting piles. Optimize program running energy consumption. The serial program CEOM is implemented based on GCC compiler. The experimental results show that according to different time constraints and program CEOM, the maximum energy saving of the system can be obtained by 12%.
【学位授予单位】：国防科学技术大学
【学位级别】：硕士
【学位授予年份】：2012
【分类号】：TP38

【参考文献】