High-performance Computing System Memory Subsystem Performance Prediction Model

In recent years, the high performance computing technologies becomes very popular in the manufacturing and researching fields. Meanwhile, the performance measurement and evaluation, especially the performance predication of High Performance Computing systems have become more and more important. Performance models are widely used for performance prediction. Using a performance model, one only needs to get some information on applications and machines and do some calculation to get the predicated time. It is more efficiently and precise compared to other methods of performance predication.However, some existing performance models are closely related to the numeric algorithms used in the applications, and would take much time and need many experts to do so; some other models and tools are based on some kind of computer architecture, and are only available to the researchers' strategic league.This paper stresses on the memory system performance model of the HPC system. It analyzes a framework of model which uses the machine profile and application signature (in loops) to predicate the performance, and transplant the modeling tools to the Intel X86 architecture, only considering loops of the applications, which can be used to simplify the process and reduce total time of model calculation. It also analyzes the problem that the application signatures obtained by Multi-MAPS cannot cover all of the memory access patterns of HPC applications and give out some suggestion on how to solve this question.The experiment results show that we can get predication results with expected precision using the performance model which has been migrated to Intel's X86 architecture.