Research On Automatic Parallization Of The CFD Programs With Time Sche(?)(?)es

With the development of parallel processing technology and the extended application of parallel computers,parallel computers are now in great demand. However,parallel program design is still difficult for most common users. To solve the problem,automatic parallelization of sequential programs has been studied since 1980s,which has been made great progress in these years. But application results are still disappointing. There is still no widely accepted and authorized Automatic parallelizer. Because Computational Fluid Dynamics (CFD) is widely used in the field of large scale science and engineering computation,and explicit difference computing method is widely used in CFD,we made deep research on the automatic parallelization of explicit difference programs for CFD.We analysed the traditional automatic parallelization technology,including dependency analysis theory,program transformation technology,parallel scheme and the optimization of related synchronization and communication etc,which are the theoretical basis of the whole article. CFD computing features,especially the features of explicit difference computing,have also been further ananlysed. We also summarized drawbacks of traditional automatic parallelization technology used in CFD:small parallel granularity,difficulty in attaining global identical data partition,and difficulty in attaining high parallel efficiency on distributed memory system.We deeply analyse the automatic parallelizer NPUPAR,which is developed by parallel processing laboratory of Northwestern Polytechnical University. And we improved it effectively. NPUPAR is a source-to-source automatic parallelization system,which can automatically parallelize sequential FORTRAN77 program into parallel program based on PVM system. NPUPAR can guarantee the correctness of iteratively field computing problems. But due to the existence of mass complex structure in sequential program and the diversity of the programming style,conservative analysis is generally adopted in NPUPAR,which leads to low parallel efficiency. We summarized several main reasons of low parallel efficiency of the NPUPAR system,and improved it accordingly. In our test,the efficiency of the programs generated by modified NPUPAR increased remarkablely (about 10%). We also make deep quantitative analysis for the super- linear speedup in the test.Cooperating with researchers in CFD,we present an automatic parallelization method based on application-understanding and extend the research from CFD to the field-dispersion computation problem. This method describes the essence of the field-dispersion computation from the point of application. And we establishes Domain-decomposition Parallel Computation Model from object-oriented viewpoint. The method looks on every sub-domain generated by domain decomposition as a single object. For interior boundary introduced bydomain decomposition,we use domain-overlap technique by message passing of interior boundary between objects. Several real problems are simulated,which proves the validity and feasibility of the proposed method. The method provides key technology support for the implimentation of automatic parallelization system in the next stage.