The Research Of Program Automatically Parallel Auxiliary System Based On SUIF Platform

The amount of computation of various disciplines is increasing with the continuingdevelopment of science and technology, in a basic computational program, serial executionalways takes hours or days for computation, which makes the research and development ofparallel computing to be more significant. To build a high-performance computing platformwhose computational speed is500,000times/per second and the storage is100TB, theComputing Center of Jilin Province update equipment’s continually. However, usually, theuser of high-performance computing center is not researchers of Computer Science, and hencethe user can only perform their own algorithms and serial computation instead of writingprograms which can perform parallel computing, this phenomenon delayed the researchschedule and wasted the useful research resources in Computing Center. Therefore, theautomatic parallelism of program developed rapidly.SUIF is a program analyzing platform developed by Stanford University，it cantransform serial program programmed by C into parallel program of C, but parallel programof C has no great expansibility, and the program has to be changed in order to adapt todifferent number of cores under the condition of existing different computational nodes. Thereare130blade nodes and8fat nodes in the Computing Center of Jilin Province, according todifferent programs, different numbers of computing nodes will be distributed. According tothis situation, this paper promoted an automatic parallelization system based on sharedmemory computer, which can use serial C program as input to generate parallel programwhich includes OpenMP parallel compiling instruction sentences via analyzing and handleprogram.OpenMp owns favorable extensibility and portability and can be self-adapt to differentnumbers of cores for its unified norms, unlimited from the API of operating systems, and theability to create and adjust the number of threads of processors dynamically.Based on the situation of the Computing Center of Jilin Province and SUIF programanalysis platform, this paper put forward an automatic parallelization system. Major works ofthis paper are as follows 1Researched the technology of automatic parallelization, includes programintermediate representation, SUIF Abstract Syntax Tree, OpenMP parallel compilinginstruction sentences, and the analysis of the program relevance. Researched the testmethods of program relevance in deep, which includes generalized GCD, boundarytesting, inequality testing and accurate testing.2Conducted research on the efficiency of program parallelism, analyzed the additionaloverhead which affects the acceleration ratio of program paralleling, the additionaloverhead includes memory accessing, load balancing and parallel control.3Designed the system according to the goals of it, on the basis of the intermediaterepresentation and relevance analysis of source program of SUIF, to get the analysisresult, parallel processed to result and inserted OpenMP parallel compilinginstruction sentences, obtained OpenMP parallel program finally.4Realized the system according to the design, based on the accuracy and accelerateratio of the experimental verification program, and by compared with the C parallelprogram, the OpenMP parallel program generated by this system has favorableexpansibility and portability.