Research Of Efficient Computing Method On GPU-based LS-SVM

As a machine learning algorithm, LS-SVM (Least Square Support VectorMachine), which has many advantages such as simple model, high training efficiencyand learning ability, has been widely applied to many research fields including timeseries forecasting and fault diagnosis and has a good prospect. However, thecomputing complexity of LS-SVM algorithm is huge, the execution time of the LS-SVM will grow rapidly with the input data sample amount. This becomes thelimitation of the LS-SVM application. Therefore, research of efficient computingmethod on LS-SVM has essential significance. GPU computing, a newly developingtechnology, could realize high performance computing in a lowcost and small sizeway. Thus, the study of the GPU based efficient computing on LS-SVM algorithmhas an important research value.This thesis analyzes the research status of efficient computing method on GPU.Then we build the GPU development system using CUDA (Compute Unified DeviceArchitecture). After that we analyze the computation flow of LS-SVM algorithm.Based on this analysis, we designed the computation structure and implement the LS-SVM algorithm on GPU. We also performed comparison research and optimizationstudy.Fisrt of all, we combine CUDA with Visual Studio software to build the GPUdevelopment platform, and adopt the Parallel Nsight to construct the debugplatform. Then, we implement the LS-SVM algorithm in CUDA C programminglanguage.After the implementation of LS-SVM on GPU, we design the comparisonexperiments on different platforms to perform the comparative study using mobilecommunication traffic data. The results show that the GPU computing has very highcomputational efficiency. Simultaneously, we also study the optimization method toimprove the performance of LS-SVM on CUDA. Finally, we test the performanceof optimal implementation through experiments.The experiments results show that the CUDA based LS-SVM could gain highperformance compared to PC platform. And the optimal implementation of LS-SVM performs even better than FPGA based reconfigurable platform, which couldprove the validity of optimal method of CUDA. This optimization research could also provide reference to other CUDA applications.