Research On Directive-based Parallel Language For Sunway Taihulight Supercomputer And Design Of The Compiling Optimization

At present,heterogeneous many-core architectures have become the first choice for building supercomputers.However,the many-core architecture has brought new and higher requirements to applications while bringing significant performance improvements.In particular,because the intergenerational development of the system architecture and the development of the application have been disconnected before,the many-core architecture faces difficulties in application migration,development and optimization.Sunway Taihu Light supercomputer is all built by home-made SW26010 many-core processor,and the problem of application migration and optimization is more serious.How to design a compatible parallel programming language for home-made many-core architecture,and build the optimized compiler system,to make the Taihu Light System more versatile,and thus exerts greater application benefits.This paper takes Sunway 26010 many-core processor and Sunway Taihu Light supercomputer as the main research object and optimization platform.It mainly studies the implicit parallel language design and the compiler optimization techniques that support the efficient porting and development of applications in the many-core architecture with master-slave shared memory.Specifically,the main work and technical innovations of this paper include the following three aspects:1?The Open ACC* language design for heterogeneous many-core processor architecture is proposed.This paper proposes a storage abstraction model for heterogeneous many-core processor architecture by analyzing the differences between the memory models of mainstream core architectures,and the implementation of Open ACC standard language text on the Shenwei 26010 many-core processor.At the same time,based on the storage abstract model,a series of language function design is proposed to utilizing and describing the high-speed memory inside heterogeneous many-core processors.It provides a full set of language functions for describing and utilizing on-chip private memory,optimizing many-core data transmission,and mining heterogeneous fusion many-core architecture features.2?A compiler architecture design for Sunway Taihu Light system is proposed.In detail,the architecture consists of major modules such as heterogeneous fusion compilers,accelerated thread support libraries,and heterogeneous runtime loaders.It can achieve heterogeneous fusion optimization at multiple levels such as compiling and linking when dealing with different instruction and structural characteristics between the main core and the slave core.This paper presents some implementation technology of Open ACC*,such as data distribution analysis technology based on affine analysis and data distribution processing technology based on heterogeneous cooperation,with provides a strong support for the efficient execution of Open ACC* applications.3?A series of compiler optimization techniques.Specifically,through analysis of the main contradictions and optimization difficulties in the optimization of the Sunway 26010many-core processor,such as structural differences between the main core and the slave core,rich storage hierarchy,streamlined structure of the slave core,and scarcity of on-chip local storage,this paper studies and proposes a series of optimization techniques for the main core memory structure of Sunway 26010 many-core processor,including memory access compilation optimization technology,compiler optimization technology for heterogeneous many core structure features,and multi-mode memory access optimization technology for Open ACC*.These technologies provide an effective optimization method for improving the performance of programs on the Sunway Taihu Light system.Based on the research results of this paper,two practical applications of CAM-SE and SWLBM,and 15 topics in SPEC ACCEL V1.0 were used to verify the effects of application transplantation and optimization in the Sunway Taihu Light supercomputer.The experimental results show that the Open ACC* programming language,compiler design and compiler optimization techniques proposed in this paper are correct and effective.They can meet the needs of many-core programming,porting and optimization for most applications,and can support the efficient operation of the application on the Sunway Taihu Light system.