Compiler Autotuning Based On Domestic Mutil-processer

Over the past two decades,the peak performance of high-performance computing has continued to increase,achieving a leap from T-level to P-level and then to E-level(Exascle Computing),providing humankind with a steady stream of basic computing power support.As the processor technology is gradually approaching the physical limit,the single-core performance of the processor slows down.In order to meet the increasing demand for computing power of human beings,various acceleration components are emerging one after another.The processor architecture presents the characteristics of isomerization and fusion.People need to be familiar with the hardware architecture to realize the efficient mapping from the top-level algorithm to the bottom-level hardware.With the increasing complexity of the processor architecture,the programming wall problem has become more and more serious.Application development,transplantation and optimization are facing more severe challenges.Modern compilers integrate hundreds of compilation optimization techniques.Through compilation and optimization,the code can be automatically transformed.This provides important help for people to perform code tuning.However,the combination space of compilation optimization is huge,and the interrelationship between compilation optimization is complicated.In the process of algorithm-to-hardware mapping,it is difficult for people to make enabling decisions for compilation and optimization.Mainstream compilers use a coarse-grained fixed compilation optimization level to assist users in using compilation optimization techniques.However,this optimization level strategy ignores the details of hardware features and code logic,and cannot meet the needs of high-performance applications.In recent years,compiler adaptive tuning has become an important technology for compilation optimization management.With iterative feedback as the main means,the adaptive tuning of the compiler can assist users to enable compilation optimization in a fine-grained form.Based on the adaptive tuning technology of mainstream compilers,this paper conducts a research on the key technologies of fast adaptive tuning of compilers.The main contributions of this article are as follows:(1)We propose distributed parallel iterative tuningIn this paper,distributed parallel iterative tuning is proposed to address the problem of high time overhead of traditional iterative tuning.The distributed parallel iterative tuning proposes various means of acceleration: based on a multi-core environment,the iterative process is distributed to multiple computational nodes for parallel acceleration;based on cross-compilation,the compilation and execution of the program is distributed to different nodes for performance,constructing a pipeline for parallel acceleration;based on the assumption that compilation optimizations are independent of each other,the reduction process is distributed to multiple nodes for parallel acceleration.The parameter configuration interface provided by parallel iteration allows the model to switch between parallel iterative tuning and serial iterative tuning.The parallel iterative tuning introduces an efficient runtime library to ensure the stable operation of the tuning process.(2)We innovatively propose a compilation optimization prediction model based on data analysisThis paper addresses the problem that parallel iterative tuning is not applicable in the field of high performance computing where programs are mostly run for long periods of time,and proposes a compilation optimization prediction based on data analysis.To address the problem of the lack of a generic dataset,the prediction model relies on the random code generation technology,and proposes a code-structure oriented dataset construction method.The model is based on the assumption of compilation optimization independence,and the optimal number of iterations is inferred by considering the time overhead and tuning effect of iterative tuning.Based on the constructed dataset,the prediction model uses the KNN algorithm to make compilation-optimized predictions for new programs.The prediction model addresses the problem of prediction accuracy and proposes an accuracy enhancement method with redundant outputs.The proposed model is highly portable and the system can be quickly reconstructed in other systems based on the proposed dataset construction method.Overall,the prediction model is based on the static features of the program and therefore the tuning overhead is significantly decreased,which makes the proposed prediction model a practical solution in the engineering field.(3)We deploy and validate the proposed distributed parallel iterative tuning and dataanalysis-based compilation-optimization prediction prototypes on the Shenwei Taihu Light domestic high-performance computerThis paper implements the proposed distributed parallel iterative tuning and data analysisbased compilation-optimization prediction based on a domestic high-performance computer.Through experiments,the parallel iterative tuning model is tested for iterative synchronous parallelism,pipelined parallelism and reduction synchronous parallelism.The parallel iterative model can achieve acceleration of the iterative tuning process on the basis of guaranteed program optimization effect.The results show that the prediction model not only provides better optimization results than CK,but also reduces the overhead of program optimization.Based on the-O3 compilation level,the core codes of the SPEC CPU 2000 benchmark suite and the core codes of the SPEC CPU 2006 benchmark suite were optimized and predicted,and the performance was improved by an average of 14% and an average of 25%,respectively.Based on the-O3 compilation level,the manual optimization of the actual subject of Shengli Oilfield was performed for tuning and testing,and a performance improvement of 15% was obtained.