Study Of The Dynamic Voltage Scaling Technique And Implementation Of Energy Optimization Algorithms For High-performance Computer Systems

Improving the efficiency of high-performance computer systems is one of the mostconcerned issues in the high-performance computing field. The higher energy thesystems consume, more huge operation cost is. Power consumption also increases theproduction cost of chips, and the advanced techniques are needed to cool the systems.The increasing power consumption leads to high running temperature, which reducessystem reliability and stability. So low power techniques are critical for the problem ofpower consumption.Dynamic voltage scaling (DVS) is an effective low power technique supported bycurrent most CPUs. It reduces CPU energy consumption through dynamically loweringthe CPU supply voltage and clock frequency. But the reduction of energy consumptionoften goes with the prolonged execution time. The longer the system runs, the moreenergy other computer components except CPU will consume. As a result, the systemenergy consumption does not necessarily be reduced during program execution period.In the thesis, we focus on how to reduce system energy consumption by using DVS. Forhigh-performance computers, since it is complex to maintain large scale computersystems, high-performance computing systems generally do not be shut down or intohibernation after the completion of tasks, but be transferred to idle state. As a result,considerable energy is consumed in idle state.In order to compare the energy saving effect of DVS, according to characteristicsof idle time for high-performance computing systems, in the thesis we present thesystem energy model (SEM) considering the effect of energy consumption of idle state.The model not only considers energy consumption within the program execution time,but considers the energy consumption of the system within the same time, whichincluding the energy consumption of idle state. We test the effect of DVS in threecomputer systems. The experimental results show that, using SEM is moreadvantageous to evaluate the energy-saving effect of DVS technique inhigh-performance systems.Based on SEM, We put forward energy optimization model (EOM) constrained bya time deadline. In the model, we assume that the whole program consists of a set ofprogram regions, and each region’s operating time and energy consumption underdifferent frequency is known. The model selects a region to run at a lower frequency soas to optimize the system energy consumption while performance penalty beingconstrained by a specified deadline.When the whole program being considered as the only region in the EOM model,this paper presents a simplified energy optimization model (SEOM).The experimental results demonstrate the effectiveness of SEOM for different computer systems. Themaximum system energy saving is up to16%for parallel programs.When function structures and outer loop structures in the program beingconsidered as the candidate region set in the EOM model, we approximately convertEOM model into a complicated energy optimization model (CEOM). CEOM measuresexecution time of all regions under different operating frequency and inserts DVS codesby means of instrumentation. We give a realization of CEOM for serial programsbased on GCC compiler. The experimental results show that under different timeconstraints, CEOM can obtain12%the largest reduction of system energy.