Edited by 颍川散人 Data: 2022.2.23
  • KB(103)和KIB(210),同样有M,G,T等

  • personal mobile device (PMD):电池,联网,贵(卡脖子在电池

  • Cloud Computing:

    Warehouse Scale Computers (WSCs)(大规模云计算,集群,阿里云百度云等,耗能大)

    Software as a Service (SaaS)(平台无关的程序,底层虚拟机,负载平衡,任务可拆分)

  • 机器语言,软硬件接口,什么决定程序的性能

  • performance & cost (最少的cost做出最高的性能)

  • Performance(后两项为本课程的任务)

    • Algorithm
    • Programming language, compiler, and architecture(语言,编译器,体系结构)
    • Processor and memory system(处理器,内存)
    • I/O system (hardware and operating system)
  • Eight Great Ideas in Computer Architecture

    • Design for Moore’s Law(摩尔定律)
    • Use Abstraction to Simplify Design(抽象简化设计)
    • Make the Common Case Fast(常用的跑得快一点)
    • Performance via Parallelism(并行)
    • Performance via Pipelining(流水线)
    • Performance via Prediction(预测)
    • Hierarchy of Memories(内外存层次化设计)
    • Dependability via Redundancy(可靠性,取舍可靠性和性能)
  • Levels of Program Code(High-level language--Assembly language--Hardware representation从高级语言到底层)

  • 冯诺伊曼架构

  • 触摸屏(输入输出合二为一)

    • multiple touch(并联数据通路,可以一起编码
  • LCD screen(frame buffer)

  • 存储

    • volatile(断电数据丢失--内存)
    • nonvolatile memory(外存)(磁盘(易于恢复 ),flash盘(不能恢复数据),光盘(长期保存))
  • 网络

  • Defining performance: response time(反应时间),throughput(吞吐量--单位时间能做多少任务)

  • 性能主要看cpu时间,并行时考虑吞吐量

  • relative performance

  • measuring executing time(主要算cpu时间)

  • CPU time :周期数乘以周期时间(以高频率,主频越高越好)clock time * clock cycles

  • CPU Time = (instruction Count * CPI)/Clock Rate(指令数乘以完成每个指令需要的周期数)

  • CPI = 周期数 / 指令数 (平均每个指令的周期数)

  • Power C =0.5* apacitiveload * Voltage * Frequency switched

  • Multiprocessors(片内并行,降低主频,重点在于装载平衡,优化同步通讯)

  • CPU Benchmark(CPU 跑分)(找到CPU的缺陷)(算性能算power)

    • SPEC CPU Benchmark
    • SPEC POWER Benchmark
  • Amdahl’s Law(改进了多少)改进后 = 改进部分/改进倍数 + 未改进部分

  • Concluding Remarks