Research On Performance Modeling Technique For Large Scale Scientific Computing Applications

As large-scale Scientific Computing Applications(SCAs)are the major applications that occupy the most running time of High Performance Computing(HPC)systems,the computing efficiency of these applications has long been a research focus.However,the gap between actual and the expected performance is increasingly widened due to the growing complexities of both machines and scientific applications.To bridge the gap,performance analysis has been considered as a necessary step,and performance analysis tools are becoming one of the most critical component in today's HPC systems.Performance modeling,the core technology to identify key performance characteristics and predict potential performance bottlenecks,is becoming an indispensable tool to understand the performance behaviors and guide performance optimizations of HPC applications.Meanwhile,numerous challenges and opportunities are introduced by the complexity and enormous code legacy of SCA's,the diversity of HPC architectures,and the nonlinearity of interactions between SCAs and HPC systems.This research focuses on how to develop an efficient performance model,considered as the next generation of performance analysis,to characterize the performance behaviors for real-world large scale SCAs.The main contributions of this study are demonstrated as follows:· Propose and establish an analytical performance model for the complicated Earth System Model,including identifying key kernels in each component by analyzing the algorithms,quantifying the relationships between crucial performance factors and key kernels' run-time,and building the model framework by analyzing coupled relationships among all components.The proposed analytical performance model is verified on a well-known scientific application,the Community Earth System Model,of the space of which is more than 1.5 million lines of code.The average model error of the proposed analytical performance model is about 10% over general-purpose multi-core platforms.· Propose an application-independent Resource-based Modeling Alongside Time(RMAT)method,and implement an automatic performance modeling tool.Different from the traditional instrument method,RMAT takes the advantage of the hardware counter-assisted profiling to collect performance data and quantify the relationship between crucial performance events and detected kernels' run time.RMAT can provide high-quality descriptions on the characteristics of nonlinear performance behaviors between applications and computing platforms,and can also provide accurate predictions to the potential performance bottlenecks.Furthermore,RMAT can limit the average model error around 8% on both multi-core and many-core platforms.· Propose an efficient process layout optimization method for multi-physics coupling applications.By introducing the concept of rectangular packing,the proposed process layout optimization method decomposes the searching process into two steps: eliminating the repetitive layout and searching the optimal process layout.Both reusing the optimal sub-layouts and pruning the search space according to parallel performance analysis are introduced to further decrease the searching overhead.Meanwhile,by combining the proposed performance model with the process layout scheduling,an automatic process layout scheduling tool is developed and has been successfully used in several earth system model projects.