On Parallel System Performance Analysis Based On TCPN

High performance computing is widely used in science and engineering to solve large-scale computation problems. However, with the scale expanding of high performance computers and the peak performance rising of hardware system, the sustained performance achieved by the real applications does not increase in the same scale as the peak performance does. The gap between them is widening. As a result, how to find the bottleneck of the system, guide the optimization of the system design and improve performance have been the most important problems that should be solved in high performance computing study.Aiming at predicting the performance of the parallel applications quickly, we research on the performance analysis model of parallel system and the key technologies. The theories on this topic, such as the determination of the key performance influence factors, performance modeling and verification of the parallel system, the improvement of the parallel application developing method, formal analysis method of the system performance are studied. The actual practice of these theories is carried out in the project-'the study and implementation of aircrafts RCS exact parallel computing'. The main conclusions of this paper are as following.(1) A novel method oriented performance of parallel programming is put forward.The performance analysis is realized by stages, which means the performance engineering should run through the whole development process of program. It will change the old development mode of coding-before-analyzing and reduce the overhead which is caused by finding the logic error and performance bottleneck too late.(2) Performance modeling scheme and verification methods based on Timed Coloured Petri Net are proposed.The performance modeling scheme which is based on the PRM theory and in-level finite TCPN is put forward. And all factors that can influence the performance of the parallel system with two aspects are considered. The experiment method of Plackett-Burman matrix is used to extract the key performance influence factors quickly. Their properties and relationships can be determined by Principal Component Analysis method. Therefore, several independent synthetic indexes can be got from many performance influence factors. It makes the system performance model can describe the performance characteristics of parallel machine and parallel application accurately and abstractly. This is the base of the further performance analysis.A classification method based on similarity and match degree for parallel system is proposed. The distance d between the modeling objects is calculated in order to decide the parallel machines (or the parallel applications) are similar or not. It can reduce the difficulty of performance analysis, such as congener application programs run on the same machine or vice versa, by accumulation the performance evaluation results of typical parallel system. The match degree sim between the parallel machine and parallel application is computed in order to judge if the parallel machine is the better operation environment for the application program.The concepts of correct concurrent procedure parallel application and the basic properties of the corresponding performance model were discussed. Furthermore, the correctness verification algorithms of the parallel application and their performance models are put forward.(3) Performance analysis and evaluation methods of parallel system based on S-TCPN are introduced.The performance evaluation indicators of the parallel system are distinguished at different level. The macroscopic evaluation index is based on the Amdahl law and alters the contradictory phenomena that the parallel processing speed is low but the speedup ratio and the parallel efficiency are high by calculating the efficiency of parallel optimization. The microscopic evaluation index describes the performance characteristics of the program. The performance data can be obtained by the variation of token in S-TCPN.The state space compression method of the parallel system performance model is raised. Otherwise, the state space explosion will cause more difficulty for performance analysis. The CPN Tools and Meta Language are used to calculate the microscopic evaluation index. It helps to locate the performance bottleneck and optimize system.