Research On Automatic Code Generation And Optimization Technology Of Data Distribution And Collection

A parallelizer is a compiler which identifies potential parallel factors in program automatically and translates the serial programs to equivalence and high effective parallel programs. Parallelizer can be divided into two important parts: Recognition of Parallelism and Automatic parallel code generation. By using the information getting from front-end of code generation, parallel program adapting to the architecture of target machine can be generated.The structure of parallel code is very key to the efficiency of the parallel program.This thesis mainly researchs the generation technology of parallel code for distributed memory machines and the relational optimizing algorithms. Firstly, the traditional data distribution and collection algorithms and their realization are discussed. Then the accurate data collection technology is done with the information of the computation decomptation and last write relation analyzing. This technology reduces the times and data volumn of the communation, and improves the efficiency of the parallel program. Also the RPEM(Redundant Parallel Execution Model) is analyzed and impoved to remove the unnecessary data collection and simplify the Model. Based on this, the data distribution and the collection algorithms are optimized. Finally, all the algorithms and optimizing means are tested to verify their validity and efficiency. The test results show that all the algorithms are correct and efficient. The performance of the program is improved greatly with the optimization algorithms.