the conclusion chart 小结think of power7 performance - ibm.pptVIP

Optimizing your EoD/EoM processing Applications with POWER7 and IBM i tuning 如何在POWER7和IBM i上优化EoD/EoM(日终/月结)批处理应用 Acknowledgements – 致谢 Many thanks to: Helen Olson-Williams - Lab Services Performance and Scalability Services  Anil Madan - Consulting IT Specialist, Lab Services Performance  Charles Farrell - Lab Services Performance and Scalability Services Mark Funk - Software Performance Analyst, Systems HW Development Agenda – 议程 POWER7 Architecture and Performance Features – Power7架构及性能特点 Application Performance Characteristics – 应用的性能特征 Performance data collection and analysis – 性能数据的收集与分析 Tuning applications and system for best performance – 应用和系统的性能调优 Hardware configuration optimization – 硬件设置的优化 Comparative Anatomy … POWER6 vs. POWER7剖析 Important points: “Multi-core” is not new. Fewer chips for same core count. POWER7 has 4x more cores per chip Cache topology is very different. The latencies have improved with POWER7. L3 data laterally cast out into neighbor L3. POWER7 has fewer memory DIMMs per core. Systems Made From Multiple Chips – 系统由多芯片构成 Simultaneous Multi-Threading (SMT) – POWER Systems 并发多线程 4-WAY SMT (Simultaneous Multi-Threading) – 四路并发多线程POWER5/6 have 2-Way SMT Cache Design Enhanced… Lateral Cast-out缓存设计增强…横向换出(In preparation for explaining “TurboCore”)  “TurboCore” Application Performance Characteristics 应用性能特征 Batch performance characteristics; single vs multithread 批处理作业性能特征：单线程 vs 多线程 Multithreaded applications can automatically take advantage of new POWER features Typically I/O bound Single threaded applications will need some tuning Typically uses only a single core at a time (e.g. max 6% of a sixteen way) Typically I/O bound CPU or disk I/O bound? CPU限制，还是硬盘I/O限制？ Low overall CPU utilization, high disk utilization, lots of memory paging Tuning techniques available to make an I/O bound job CPU bound Batch jobs waiting for something else? 批处理作业在等待(waiting)什么？ Wait for I/Os, e.g. disk configuration #