自动寻找SPMD模型公共变量的一种方法

A Method of Searching for Common Variables of SPMD Model Automatically

自动寻找ＳＰＭＤ模型公共变量的一种方法肖骊，康继昌关键词ＣＶＤ（化学气相沉积），预涂层，石墨粉，金刚石膜中图分类号ＴＢ４３并行软件的研制是当今并行处理领域的热点和难点之一，程序自动并行化是其中的一个重要领域。但是完全的自动并行化实现起来非常困难，作者...

SPED (Single Program Multiple Data) programs find wide use in aeronautics such as CFD (Computational Fluid Dynamics). In about five years of research on parallel computation, we at Northwestern Polytechnical University have not yet found in the open literaturefUlly satisfactory information concernins automatic transformation of serial program into parrallel program for SPMD. Now in this summary, we present our research results on perfecting such transformation.In order to perfect automatic transformation, we abandon the usually adopted unified memory space or virtual shared memory. Instead, we let memory of each node, which can prOCess data individually, store its own share of data. However there is the need for each node to make use of data stored in other nodes. Such mutual needs must be clearly defined,and we use the term 'common variables' to mean such mutual needs. When such common variables are located, or clearly defined, the use of my research results not given in this summary on automatic communication and automatic sychronization between nodes then allows parallel computation to be implemented. Thus, the main contribution of this summary is how to locate or search for such common variables automatically.The main ideas of searching for common variables in our research are:(1) As the characteristic of SPMD models is to process large data set, often defined asarrays, by same program simultaneously, the data is distributed evenly into nodes in a certain direction of multi-dimensional space.(2) Two kinds of distributions of common variables require different methods for location of common variables' For common variables distributed on the border of each block, weput forword a new algorlthm (see Chinese text) which finds all such variables with high efficiency. This kind of distribution is very common. For common variables distributed randomly, we find them through enumeration, whose low efficiency is not serious as this kind of distribution is quite uncommon.