Research On Computing Feature Classification Of Complex Computer Architecture

At present, the scale of computer system is growing gradually. Users demand thatthe performance of computers improve continuously which also contribute to aconcurrent execution of a large number of heterogeneous applications and threads.Therefore, making a global analysis about the computing features of the complexcomputer architecture has become the focus of research. Most of the traditionalmethods of analysis use detailed software simulation and other related technologies,in order to improve the speed of simulation, the researchers also proposed someimproved simulation, such as statistical simulation, sampling simulation andaccelerated simulation, which provides a more efficient method for the analysis ofcomplex computer architecture. However, there are some limitations in computingfeature analysis of large-scale heterogeneous computer architecture.This thesis applied classification methods based on machine learning in featureanalysis, which can guide the design of computer architecture better, in order toanalyze and use features of complex computer architecture more comprehensive andmore effectively. The main contents of this thesis are as follows:1) This thesis discussed related analysis technologies of complex computerarchitecture. Analysis and comparisons were made between the traditional clockcycle simulation techniques, statistical simulation method, sampling analogtechnology and accelerated simulation, then discussed the pros and cons and theapplication prospect of features analysis in computer architecture based onmachine learning.2) A computing features extraction method of complex computer architectures wasproposed in this thesis. The computing features extraction of architecture providesan important basis for the classification. Mainly the principal component analysiswas applied to extract the computing features of complex computer architecture,after the computing features IPC (number of instructions executed per clock cycle)and power extracted the thesis described the obtained feature space withassociated methods of mathematical statistics. Studies have shown that effectiveextraction of computing features in complex computer architecture can help toimprove the accuracy of classification of computing features.3) This thesis realized a computing feature classification method based on Boostingdetermination model. Training was made on the samples based on Boosting algorithm using superimposed weights. These samples contain extractedcomputing features IPC and power and corresponding combination of multi-coreconfiguration parameters. AdaBoost classification model was achieved withcontinuous iterative training in the thesis, which will be used for classification oftwo computing features of complex computer architecture. In the experiment,compared our results to the previous studies of BP and RBF neural networkclassification model, the mean error of determination model is about3to4percentwhile the mean prediction error of neural network is about4to5percent. Thusthis model is better than the neural network model with higher accuracy, and themethod proposed can be applied in computing feature division of complexarchitecture better.