The Research Of Parallelization Technique Based On Shared Memory Structure And Optimization

Some parallelization techniques for parallelizing sequential programs are proposed bothat home and abroad. However, because of the complexity and overhead of parallelization, theefficiency of parallel execution may be inferior to that of sequential execution for the samesequential program. In order to decrease overhead and improve parallel execution efficiency,the research of optimizing some parallelization techniques is necessary.The CorD (Copy or Discard) model uses the speculative parallelization technique todevelop parallelism of sequential loops especially sequential loops carried data dependencebetween iterations on multi-core processors, and automatically generates parallel programswith OpenMP compiler directives. But for loops with misspeculation occurring earlier, theefficiency of speculative parallelization technique is very low. Therefore the paper proposes amethod named loop pre-computation, replacing the speculative parallelization technique toimprove the CorD model.The realization of the loop pre-computation method first needs to analyze the code of asource program to recognize its loop body. Then analyze data dependence of the loop torecognize the active variables resulting in data dependence between iterations and theirdependent variables, and put the statements producing these variables in order so as togenerate the pre-computation slice of the loop. Finally, insert the pre-computation slice of theloop into start of the loop body, and generate parallel programs with OpenMP compilerdirectives.The paper parallelizes some loops carried data dependence between iterations using theoriginal CorD model and the improved CorD model. The experiment results show that theimproved CorD model is more efficient than the original CorD model. And the improvedCorD model is suitable for loops with misspeculation occurring at any time.