On The Automatic Generation Of GPGPU Power Stressor Based On Fine-grained Instruction Combinations

With the advent of the era of big data and artificial intelligence,the demand for peo-ple to use graphics processors(GPU)for general-purpose computing has surged.GPU has evolved into a graphics processor suitable for general-purpose computing(GPGPU).At the same time,GPGPU has been playing an important role in the current HPC land-scape.In the latest top500 supercomputer list in June 2020,333 supercomputers contain GPU,accounting for two-thirds.By 2023,Gartner predicts that 10%servers will use GPGPU to accelerate tasks.However,it is a big problem to design the power supply and heat-removal system of GPGPU.Most of current designs are conservative strategies based on experience,leading to over-design problems in the power supply and heat-removal system of the GPGPU,which ultimately wastes a lot of energy.Some power supply and cooling sys-tem of GPGPU have under-designed problems,which sometimes causes GPGPU to fail to operate normally.The GPGPU maximum power consumption test program can help people accurately design power supply and heat-removal systems.Meanwhile,GPGPU has a large number of parallel computing cores,multi-level storage systems and com-plex internal interconnection structure,which makes it very difficult to directly write the maximum power consumption test programs of GPGPU.At present,there are the following problems in the research of power consumption of GPGPU:1.The existing power consumption test methods are complicated and cannot quickly test the maximum power consumption of GPGPU of different architectures;2.The power consumption of GPGPU tested by the existing methods is not high;3.The existing methods do not com-prehensively consider the relationship between load program characteristics,hardware performance counters,various functional components and power consumption.Aimed at the above problems,this thesis proposes a maximum power consumption programs framework of GPGPU based on fine-grained instruction combinations.The key of the method in this thesis:1.Design power consumption program generation algo-rithm based on fine-grained instruction combinations;2.Use a two-stage optimization method to determine the optimal value of fine-grained parameters,and then determine the optimal value of coarse-grained functional components value.Run 28 benchmark test programs on GPGPU of two typical architectures,and establish a performance analysis model.Analyze the relationship between program performance counters,GPGPU com-ponents and power consumption,and find the importance of each factor.According to the importance of each factor,design the automatic generation algorithm of the power consumption test program.At last,study the parameter value range of the power genera-tion program algorithm through experimentsThe final results of experiment show that the dynamic power consumption of the existing method on GTX480 and Quadro FX5600 are 193.7W and 163.7W,respec-tively.The maximum dynamic power consumption obtained in this paper is 223.9W and 235.8W,respectively,which are 15.59%and 44.04%higher,respectively.Combining with static power consumption,compared with the official thermal design power con-sumption.The official thermal design power consumption of GTX480 and Quadro FX5600 are 250W and 175W,respectively.The maximum power consumption of the generated program are 265.8W and 257.1W,respectively,which exceed the official ther-mal design power.Consumption 6.32%and 46.9%.The official thermal design power consumption is much lower than the limit power consumption of the generated program,and the measured maximum power consumption will effectively guide the design of the power supply system,heat dissipation system of the GPGPU.