Parallel Transformation Of Serial-program Under Multi-core Platform

When the performance of CPU was upgraded to a high-point, single-core has been replaced in current computer system by multi-core. In order to make full use of powerful computing resources in the multi-core platform with it we can run these time-tested good serial-program more efficiently, the research of serial-program parallelism transformation has become quite important and active at now, so that we can get the coordination in the development of software and hardware. This paper focuses on the analysis of control dependency and data dependency which is the most important techniques in serial-program parallelization. According to this we propose a serial-program parallelization scheme- ‘A serial-program parallelization which based on weighted set’. O ur research focus on the analysis method of control dependency and data dependency, then for the scheme(a serial-program parallelization which based on weighted set) we have an in-depth analysis and discussion, finally we validate the scheme through the parallel transformation of t he C RAES-system which come from our actual work. The main research content is divided into four parts.(1) Summarized the main knowledge of serial-program parallelization(serial program, parallel program, parallel program design, parallel programming model and program control dependency and data dependency).(2) Explained the requirement analysis and the summary design of the scheme proposed by this paper which based on the parallel transformation of the CRAES-system.(3) Elaborated the analysis method of control depe ndency and data dependency, and gave the improved algorithm; Analyzed the possibility of serial-program parallel transformation from the three point of view: data partition, task partition and loop- level partition; Described the algorithm design and imple mentation of this paper’s serial-program parallelization scheme; Gave a debugging and optimization method for the parallel program.(4) Introduced the CRAES-system and analyzed its algorithm, made a parallel transformation for it, finally compared the performa nce of the program before and after the parallelization through experiments.