On The Relationship Between Parallel Computing Time And Storage Requirement

High performance computing (HPC) is widely used in science and engineering fields to solve large scale computing problems. The demand for promoting its performance has rapidly increased. Facing huge, complex and time-critical computing tasks, how to optimize parallel program designing to improve parallel system performance efficiency is still an important and difficult problem in high performance computing, which remains to be broken through. To solve such problem, the performance evaluation of high performance computing is the one, which should be solved firstly.It is a complicated process to develop and optimize parallel program. Time requirement and storage requirement are two important aspects that should be taken into consideration during the programming process. Refining the relationship between time and storage space is also an efficient method to optimize parallel program. In this dissertation, the relationship between time requirement and storage requirement is deeply analyzed. The theories on this topic, such as evaluation metric for time and space, the relationship between time and space, and the overhead model of parallel program and its computing processor scale's calculating method are studied. The actual practice of these theories and analyze are carried out in the project-'the study and implementation of aircrafts RCS exact parallel computing'. The main contributions of this dissertation are as follows.1. The time speedup model and space speedup model are raised.Based on the parallel program's characteristics, the speedup is adjusted, and renamed as time speedup model. In this model, the existence of the time speedup in parallel computing is discussed, and the spatial factor is placed into it. Meanwhile, the changing trends of time efficiency and computing time after parallel program optimization are analyzed.The changing regularity of storage requirement in parallel computing is analyzed, and the space speedup model is proposed. In this model, the primary characteristic of storage requirement in parallel computing is identified. And in order to obtain the parameters used in the space speedup model, two storage space statistics are proposed. One can be used to survey the total storage requirement of a parallel program, and another to census the storage requirement in a certain node.2. The relationship model between time and space and the relationship prediction method are proposed.Four kinds of time and space's relationship are analyzed, and the relative time efficiency and space efficiency diagrams are given. Using the time efficiency and space efficiency diagram, the equilibrium point to make effective use of system resources and shorten the calculating time can be found.A calculating model to use space to express time is presented. Using a relatively simple way, the processing time of the key storage space under certain parallel processor scale can be calculated with little overhead. It can be used to study the key storage space's effect on parallel program's whole performance and to provide the possibility to predict the relationship between time and space.3. The overhead model of parallel program and its processor scale's calculating method are raised.In view of the distributed memory, shared memory and distributed shared memory systems'structures, the overhead models of MPI program, OpenMP program and MPI+OpenMP program are studied. Especially, the MPI+OpenMP program's overhead model reveals the mixed program's overhead sources and the relationship between them.The characteristic of OpenMP program is analyzed, pointing out that the method to determine processor scale, when using OpenMP to develop programs, should be reconsidered. The strategies to calculate the processor scales of OpenMP program and MPI+OpenMP program are proposed, making use of the expansion condition of program size. Thus, the main difference in discussing the time and space problem under distributed memory, shared memory and distributed shared memory systems is analyzed.