Study Of Low Power Software Optimization Technology For GPUs

Lowpoweroptimizationhasbecomeanimportantresearchfieldinmoderncomputerarchitecture area. The power problem has attracted much attention from large-scaleparal-lel computing systems to desktop computing systems and to embedded systems. Firstly,the increase of power consumption means increase of computing cost. Especially forlarge-scale parallel computing systems, energy consumption increases dramatically sothat it has become an important factor that restricts the scalability of the systems. Second-ly, the increase of power consumption raises the cost of chip package and heat emission.High temperature will influence both the life and the reliability of the chip. Finally, thewide application of mobile electronic devices presents a serious challenge for the pow-er optimization of embedded processors. However, under the lead of Moore's law, thesemiconductor technology keeps improving and the density of transistor on chip is get-ting higher and higher. Nowadays more than 1 billion transistors have been integratedinto a single chip, which exasperates the power problem undoubtedly.Lowpoweroptimizationsfortraditionalprocessorshaveobtainedwideattentionandresearch. However, in this thesis, we select a new kind of processor architecture, GPU,which is very popular in recent years, to perform the low power optimization researches.GPUs(GraphicsProcessingUnits)arefirstusedtoacceleratethegraphicsprocessing,andthus have simpler architecture than contemporary CPUs. The chip resources can be usedto improve the computing performance more efficiently. Consequently GPUs typicallyrender much higher performance than CPUs. With the development of GPU's program-ming interface in recent years, people began to accelerate some non-graphics computingtasks using GPUs. A brand new research field called GPGPU (General Purpose Com-putation on GPUs) has come into existence thereby. Constructing large-scale computingsystems using CPUs and GPUs even becomes an important trend in the high performancecomputingarea. GPUhasaverypromisingprospectinhighperformancecomputingarea,desktop computing area and embedded computing area. However, GPU has much higherpower consumption than CPU because of its high density of computing units, despite itshigh power efficiency. The power consumption of GPU has become an important factorthatlimitsitsapplication. SothepoweroptimizationtechnologyforGPUisveryemergentand worth to study. Generally, power consumption technologies of processors can be categorized in-to several levels, including physical circuit level, architecture level and software level.Those methods close to the physical level can be in general use for various processors.However, they are not associated with applications. On the contrary, the methods closeto the software level can optimize the power consumption according to the characteris-tics of the applications. In this thesis, we adopt the software optimization method, whichaims at the architecture of GPU, considers the feature of GPU programs, performs theresearch oriented towards different optimization environments and objects. To be morespecific, the optimization environments include multi-GPU system, single-GPU systemand the processor in the GPU; the optimization objects include the single task running onthe GPU and the whole program consisting of multiple GPU tasks. The innovations ofthis thesis are as follows:1. Proposing a low power optimization method for CPU-GPU heterogeneous parallelsystem based on critical path analysis. According to the execution feature of GPUprograms, we represent the execution progress of a whole program including mul-tiple GPU tasks as a data structure based on graph, and thus we can identify thosetasksthatcanbeoptimizedundertheperformanceconstraintandperformthepoweroptimization.2. Proposing a static energy optimization method for multi-GPU environment basedon task distribution. Under a given performance constraint, we build the model fortask distribution in multi-GPU environment, map the GPU task onto proper GPUsand shut down the idle ones, which will lower the static energy consumption of thewhole system.3. Proposing a dynamic power optimization method for GPU based on parallelismanalysis model. Considering the processor and the memory on the GPU syntheti-cally, we study the relationship between the processor and the memory during theprogram execution, guided by a performance model based on parallelism analy-sis. Then we can optimize the dynamic power consumption under the performanceconstraint using the dynamic voltage scaling technique.4. Proposing an energy optimization method for GPU based on the software prefetch-ing technology. We deeply analyze the execution manner of GPU and establish its optimization space for software prefetching. Then, under the constraint of per-formance and power consumption, we propose the optimization methods for theenergy consumption and performance respectively.