Research On The Power Prediction Model Of Integrated GPU

The emergency of heterogeneous architecture promotes the fast development of high performance computing technology,and the heterogeneous processors have become the mainstream of high performance computing systems.However,the processor has met the bottleneck to make further development and single-core architecture cannot meet the requirements of modern applications owing to power issues.Multi-core architecture has relieved the pressure of power problems,but power is still the major concern of processors.Due to the variety of functions and on-chip resources,power issues are becoming more and more serious and put tremendous negative impacts on the advance of processors.Therefore,power is the state-of-art topic in computer architecture.The power of the processor is mainly determined by the state of processor: when the frequency and voltage are in a lower state,the power will be smaller.According to this,some people proposed DVFS which can lower frequency and voltage to reduce the consumed power of the processor without sacrificing too much performance.The power of processor in the current and other states are the necessary parameters to finish dynamic scaling,and the scaling scheme will determine the goal state according to the power parameters.This paper targets the power prediction technology of integrate GPU(iGPU)in APU,and proposes a power prediction model which is based on the linear regression method.First,we propose a power prediction model based on the linear regression method.This paper takes the characteristics of iGPU architecture into consideration,and analyses the major components of iGPU power,which includes the computing module,the local memory and the global memory power consumptions.We choose as many performance counters as possible to cover the above three parts so that the power calculation is more precise.Upon this,we propose the power prediction power model by running the benchmarks in Rodinia to collect the power profiling data and performance counter profiling data.Because of the differences between the sample period of power profiling and performance counter profiling,we cannot take the collected data directly as the input to the model.We propose the kernel extension method to extend the execution time of different kernels and try to guarantee the length of the two sample periods could match with each other,which deals with the problem successfully.After collecting enough runtime information and organizing them as the input for training purpose,we use SPSS to accomplish the linear regression analysis and finally obtain the power prediction model.The results show that the standard prediction error is just 2.12%.Furthermore,both the correlation coefficient and the adjusted R square coefficient are larger than 0.95,reflecting that the linear relationship between the independent variables and dependent variable is quite obvious,which proves the correctness of the theoretical foundation.Second,we propose two simplified power models and the approaches to build them.While applying DVFS to APU,the processor is quite sensitive to the instantaneity,which demands the power model must have low latency.What's more,due to the combination of CPU and GPU,the space for other online resources is quite limited on APU.If our power model is too complicated and includes too many performance counters,it will add extra complexity to the design and usage of APU hardware.Therefore,we propose the single-simplified and multi-simplified power model to tackle with the dilemmas.The single-simplified power model mainly explores the relationship between the type of performance counters and the standard error of power models.The results show that the class of computation performance counters is the most influential while the class of global memory performance counters is second.Then we build multi-simplified power models to explore the connection among the number of performance counters and the stability and prediction precision of power models.Our experiments suggest that when the model includes 12 performance counters or more,the stability is satisfactory and the error is less than 5%.Such power models are quite applicable.Compared to the original model,the amount of computation decreases by 20% and the latency is significantly reduced,which improves the instantaneity of the power model.To sum up,this paper targets to solve the power prediction problems in the heterogeneous processor APU and proposes the power prediction model based on linear regression method.We successfully control the error of the model.In order to meet the demand of instantaneity,this paper proposes the multi-simplified power model,which can significantly reduce the cost and latency without losing much precision.Our work is really meaningful both theoretically and practically.