Paraxle:A High-performance DSL For Big Data And Scientific Computing

In the era of big data,algorithms and applications of the field of big data,scientific computing and artificial intelligence depend not only powerful computational resources as the foundation,but also efficient and easy-to-use programming models as the medium of the interaction between the users and the hardware.Currently,programming frameworks for big data and scientific computing tend to be quite complicated to use,which is not friendly to those who are not proficient in computer science or programming,hindering the further development of interdisciplinary research and application.Domain specific language is a kind of computer programming language for specific domains;it has the characteristics of concise grammar and powerful expressiveness.Axle is a domain specific language for scientific computing,providing basic functions for big data and scientific computing.Based on the internal demand for better programming model in the era of big data,this thesis proposes Paraxle,a domain specific language for big data and scientific computing based on Axle.Paraxle has extended Axle from several aspects.Paraxle has further optimized selective parallelizable functions of Axle to increase the efficiency of these functions.Because Axle only provides limited functions about big data,Paraxle provides more big-data-related functions in the areas of regression,classification and clustering.Since Axle does not provide access for heterogeneous computing,Paraxle has implemented many functions concerning the areas of linear algebra and simulation based on OpenCL,a framework for heterogeneous computing,and encapsulated above functions into convenient interfaces,making it accessible for people with little OpenCL-ralated knowledge.Based on Akka actor framework,Paraxle has implemented selective parallel big data algorithms,which provides convenient programming interface for clustering computing on Akka clusters.This thesis introduces the implementation of Gaussian Process Regression based on Paraxle and compares this implementation with the pseudocode of Gaussian Process Regression,reflecting the simplicity and accessibility of programming with Paraxle.The experiments have been conducted to test and analyze the feasibility and efficiency of Paraxle in detail,and the experimental results support the success of Paraxle.Comparing with the original Axle,Paraxle has access to heterogeneous computing thus it can be applied into broader areas related to big data.Besides,since Paraxle is a Scala-embedded domain specific language,it inherits the powerful type system from Scala,which is helpful for further extending.