Research On Key Techniques Of Performance Models For High Performace Computing

High performance computing (HPC) is widely used in science and engineering to solve large computation problems. With the development of HPC, the scale of the high performance computers is expanded rapidly. Many new technologies and methods are introduced to improve the performance in the designing of the processor nodes. The peak performance of computers increases in a continuous and rapid way. But the sustained performance achieved by the real applications does not increase in the same scale as the peak performance does and the gap between them is widening. Performance evaluation of parallel systems, which is one of effective ways to solve this problem, can find the bottleneck of the system and guide the optimization of the system design.As the computer architectures and program structures are becoming much more complex, more and more factors may affect the performance of the programs. Furthermore, these factors interplay with each other in a complex and nonlinear way, which makes the performance evaluation of parallel systems a great challenge. Traditional performance evaluation methods cannot satisfy the need for performance evaluation of these massive parallel systems. Performance model which combines the application signatures and the machine profiles draws the attentions of the research community as well as the industry community. This method analyzes application signatures and machine profiles independently, and uses convolution methods to map an application's signature onto a machine profile to arrive at the performance prediction. Aiming at predicting the performance of the parallel applications exactly, we research on the performance model of parallel systems and key technologies.The dissertation thoroughly investigates the present status and hot points of researches on performance evaluation of parallel systems. Several important projects are analyzed, and their characteristic and short point are summarized.The dissertation considers all the factors that can influence the performance of the parallel system, and proposes a performance metric of the parallel systems on multidimensional space. The metric system defines basic performance metrics: Application Intrinsic Metrics (AIM), System Performance Functions(SPF) and System Performance Metrics Application Oriented (SPMAO) , and proposes the distance between these metrics and the similar relations among them. The performance metric on multidimensional space builds the theoretical basis of this dissertation. It set up a map from parallel systems to abstract mathematics space.Considering all the characteristic of most performance models of the most of the parallel systems, a novel performance model framework PMPS (Performance Model of Parallel Systems) based on the convolution method is proposed. This performance model has good scalability and extensibility which comes from the hierarchy convolution methods that combining the parts and the integer of the system.To decrease the dimensions of the performance metrics and reduce the complexity of the PMPS analysis model, a method named DoubleP is proposed to discover the key performance factors of the processor nodes. DoubleP can focus complex performance factors on several main components, so the analysis objects can be seen clearly. Using DoubleP, 14 key factors which can influence the performance of processor nodes and 4 main components of system's performance are found.The method to analyze programs profiles is the main means to study the application's signatures in PMPS, which is also a difficulty in the research of parallel systems performance. For analyzing the program profiles quickly, we proposed a method based on sampling. Compared with other methods, this technique can reduce the needed instruction numbers and shorten the analyze time of programs profiles on the same conditions that a certain sample error can be ensured, which means only 1%~3% instructions will be used when the error is less than 3%. Further more, a profiler named SamplePro base on sample theory is put forward and implemented in the dissertation.The performance model of processor nodes is the main part of the PMPS. The dissertation presents a performance model of processor nodes and its solving method based on regression. This method converts the performance factors and the relations between independent predicting variables, and obtains the weights of the complex and overlapped operations by determining the regression coefficients. The experiment results show the efficiency of regression method and the accuracy of the regression model, and cannot be influenced by the processor types and application signatures.The experiment results indicated that the PMPS performance model with good scalability can precisely predict the running time of all kinds of parallel applications in the parallel computers. It can also discover the performance bottlenecks of the parallel systems. The PMPS model can provide plenty of performance parameters and guiding information for designing, optimizing and upgrading the parallel computing systems.