Scheduling Algorithm For Energy Efficiency Optimization Based On GPU Cluster

With the development of GPGPU technology,GPU's computing performance is increasing continuously,and the problem of energy consumption become more and more serious,which draw people's attention to use energy efficiency to evaluate the performance of computer system.The analysis of the GPGPU energy efficiency can help GPU cluster to implement “green computing”.Nowadays,scholars have analyzed the GPGPU performance,energy consumption from the GPGPU architecture and GPGPU multiple load,moreover they building the model,or propose the GPU cluster energy efficiency optimization scheme,on the basis of this,but their works still exist some limitations.In order to better implement the GPU energy efficiency optimization.In this paper,on GPGPU feature extraction and modeling,the machine learning theory has immersed in it,and quantifies the GPGPU energy efficiency as the basis for decision making of GPU cluster energy efficiency optimization method of design.The research of this thesis is as follows:(1)Explore the factors analysis technology of GPGPU performance,energy consumption and its influence on.From the single node,analyzes the influence of GPGPU performance and energy consumption in the architecture and the load code,including SM saturation,density calculation,calculation unit number,the number of registers and the cache structure.Moreover,this paper summarizes the performance of GPGPU or energy consumption of the existing prediction methods.From the GPU cluster,analyze the cluster management technology and the existing cluster scheduling scheme.(2)Implement the GPGPU energy modeling.GPGPU energy modeling is decomposed into GPGPU performance modeling and GPGPU energy consumption modeling.In performance modeling,simulate all kinds of architecture and collecting performance counter information as the feature data through the free combination of the affecting factors of GPGPU performance.On this basis,using attribute reduction scheme to extract key features,and using a coarse-grained clustering strategy to discrete data,finally establishes the GPGPU performance prediction model based on BP neural network.The average prediction error calculation features and key features were 8% and 6% to the model.In the part of energy consumption prediction,modified the load code to simulate energy consumption in all kinds of load conditions,and extracted the key features from the performance counters as computing features.In all,establish the GPGPU energy consumption prediction model based on BP neural network which modified from a genetic algorithm to prevent the limitations.Finally,the GPGPU energy efficiency prediction model to GPU cluster was proposed as the making decision of energy efficiency scheduling algorithm,the model average forecast for two types of GPU errors are less than 1%.(3)Implement the energy efficiency optimization algorithm of GPU cluster.A GPU cluster energy efficiency optimization algorithm was designed by the improved artificial bee colony algorithm,which the GPGPU energy efficiency model is the decision-making.According to the different load and node mapping the energy efficiency and node surplus computing power as the basis for decision making,employed bees and follow bees in artificial bee colony algorithm improved the nectar find,which make the scheduling algorithm can guarantee the cluster system load balancing and energy efficiency optimization.Experimental show that compared with FCFS's own torque scheduling algorithm,this algorithm can save about 5% of the power consumption of the cluster system.The works of this thesis show in two aspects:(1)Using the BP neural network and rough set theory,the calculation feature of GPGPU is processed,which implement the data discretion and key features extracted.And,the GPGPU energy efficiency model is established.(2)Using the artificial bee colony algorithm,analyzed the GPGPU energy efficiency and the node surplus computing power.Also,the scheduling algorithm based on GPU cluster is designed to implement the load balancing and energy efficiency optimization of the cluster.