Research On Parallelization Of Machine Learning Algorithms For On-chip Heterogeneous Multi-core Systems

In modern times,with the birth of the Internet of things,the rise of big data theory,as well as the popularity of mobile Internet,we have entered an era of information explosion,and also a more intelligent era.Many large amount information processing and intelligent applications need to be implemented on intelligent device,they are all embedded systems.This situation makes machine learning to be more important,but also make the traditional embedded system to be difficult to meet the performance requirement.High performance embedded computing technology is an efficient way to solve the problem.This dissertation mainly studied the parallel optimization techniques of machine learning methods in embedded intelligent devices,relying on multi-core technology as the main means of high performance embedded computing.This dissertation analyzed the architecture of embedded multicore system,and proposed a general architecture of heterogeneous multicore system and an abstract model of heterogeneous multicore application.Then explored parallelization strategy of machine learning method and implemented in Parallella platform which achieved outstanding processing speedup performance compared with the serial machine learning method.At last this dissertation developed a heterogeneous multicore based machine learning application rapid development framework to reduce the difficult of parallel machine learning application development.The major contributions of this dissertation are stated as follows.(1)A general architecture model and modular program execution model for heterogeneous multi-core high-performance computing systems are proposed.Through the research of high performance computing system architecture,first determine the general and special CPU accelerator collaborative high performance computing is the main structure of the future high performance embedded computing,and multi-core CPU and accelerator form heterogeneous multi-core architecture based on the proposed model,after the embedded application division for multi task structure,forming a series of control unit and element combination task set,the execution model of heterogeneous multi-core architecture under the program.The execution model can effectively guide the development of embedded high-performance computing applications through task splitting mechanism and multi-core mapping method.(2)A set of parallel machine learning algorithm,parallel optimization strategy and application algorithm for heterogeneous multi-core architectures are proposed.Firstly,this dissertation researched Parallel strategies from the parallel data and model.They are data parallel based AdaBoost classification method,data parallel based rarefaction SVM training method,model parallel based multi-layer perceptron classification method and hybrid parallel convolutional Neural Network classification method.Secondly,the dissertation proposed parallel machine learning method for heterogeneous multi-core platform and its scalable technology,then studied the actual machine learning algorithms based on Parallella high performance embedded computing platform for a variety of machine learning applications implementation.(3)Aiming at heterogeneous multi-core architecture,an extensible machine learning application framework,PML-RADF,is proposed.This dissertation studied on machine learning parallel acceleration algorithm library,mechanism and communication mechanism of multi database maintenance,implementation technology research and development framework of software architecture and design,finally built the universal and scalable algorithm and rapid development framework.