第170回 第2部

Modern supercomputers are highly parallel machine combining inter-nodes process parallelism and inter-core thread parallelism. And the memory hierarchy including the cache in the node is also complicated. On the other hand, applications that run on supercomputers cannot fully utilize the performance of hardware unless high parallelization and individual node tuning are performed according to that hardware.Therefore, the two points“programming conscious of parallelism” and “programming conscious of execution performance”are essential techniques for users, researchers, and programmers who use the present supercomputers equipped with tens of thousands of processors and containing various enhancements and new functions. Here, we call the technique as the performance-optimizing techniques to application programs. Performance optimization of application is not always done by the application developer. It is difficult to interpret applications developed by others, evaluate their performance, discover problems, and solve problems. To systemize of the techniques of performance optimization will provide useful information for engineers and researchers who want to optimize the execution performance of applications. In this talk, I will talk about systemization of the techniques of performance optimization for single CPUs and high parallelism. Specifically, the following content is included.
-Classification of applications from the viewpoint of single CPU performance.
-Explain of busy time.
-Relationship of busy time and performance.
-Relationship of busy time and classification of applications.
-Relationship of busy time and performance tuning.
-Maximum performance estimation when busy time depends only on bandwidth.
-Accumulation of tuning techniques each application classification.
-Classification of problem regarding high parallelism.
-Accumulation of tuning techniques each problem classification.