tags: 計算機組織_戴碧如

Chapter 1 - Computer Abstractions and Technology

• Chapter 1 - Computer Abstractions and Technology
  • Seven Great Ideas
  • Below Your Program
  • Levels of Program Code
  • Through the Looking Glass
  • Inside the Processor (CPU)
  • Abstractions
  • Memory
  • Defining Performance
  • Responese Time and Throughput
  • Measuring Execution Time
  • CPU Time
  • Instruction Count and CPI
  • CPI in more detail
  • Performance Summary
  • Power Trends
  • Reducing Power
  • Solution of Power Wall
  • SPEC CPU Benchmark
  • SPEC Power Benchmark
  • Pitfall: Amdahl's Law
  • Fallacy: Low Power at Idle
  • Pitfal: MIPS as a Performance Metric

Seven Great Ideas

• Use abstraction to simplify design 把複雜的設計，切成一層一層的，在開發上會變得比較容易且快速。
• Make the common case fast
• Performance via parallelism 平行處理
• Performance via pipelining 接力的概念，把流程切成小部分且連續
• Performance via prediction 預處理可能的結果
• Hierarchy of memories 因為所有的資料不會一次都做使用，同一時間只會使用到一小部分的資料，所以把所有的資料放在「便宜、空間大」的儲存裝置，把要處理的資料放在「貴、空間小、速度快」的儲存裝置
• Dependability via redundancy

Below Your Program

• Application software 高階語言撰寫
• System software
  • Compiler 把高階語言轉換成 machine code
  • Operating System
• Hardware

Levels of Program Code

• High-level language
• Assembly language
• Hardware representation
  

Through the Looking Glass

LCD screen: picture elements (pixels)

• Frame buffer 紀錄 pixel 的顏色
  

Inside the Processor (CPU)

• Datapath: performs operations on data
• Control: 用來告訴 Datapath 要往哪裡走（控制信號）
• Cache memory: SRAM 小且快的 memory

Abstractions

• Instruction set architecture (ISA)
  • The hardware/software interface 軟體跟硬體之間的介面

Memory

• Volatile memory
  • RAM
• Non-volatile memory
  • Hard-Disk, Flash memory, CD\DVD

Defining Performance

要定義一個 CPU 的 Performance 不能只單看一種面向，必須多個面向一起探討。

Responese Time and Throughput

• Responese Time (execution time)
  • How long it takes to do a task 執行一個工作的總時間
• Throughput (bandwidth)
  • Total work done per unit time 每單位時間的工作量

Measuring Execution Time

• Elapsed time
  • Total response time 把所有的處理時間都考慮進去（包含I/O、OS overhead、idle time）
  • Determines system performance
• CPU (execution) time
  • Comprises user CPU time and system CPU time

一般而言，user CPU time 反映了主要邏輯程式碼的執行時間，而system CPU time 則反映了執行 I/O 操作、管理記憶體或與操作系統交互的時間。

CPU Time

• clock cycle 指的是一個時脈週期
• clock rate 則是 CPU 每秒鐘所執行的 clock cycle 數量
• CPU Time 就是只說，完成一個 task 所需要的時脈週期總數，除以 CPU 的處理速率（clock rate）
• Performance improved by （CPU Time 要變小）
  • Reducing number of clock cycles （降低一個 Task 所需的 clock cycle 數量）
  • Increasing clock rate （增加 clock 頻率）
  • Hardware designer must often trade off clock rate against cycle count

Instruction Count and CPI

• Instruction Count (IC) => 一次需要呼叫的指令數量
• CPI (Cyclesper Instruction) => 一個指令需要花費的 cycle 數量

Clock cycle 總數 = IC x CPI
CPU Time = IC x CPI / Clock Rate = IC x CPI x Clock cycle time (一個 clock cycle 所需的時間)

ISA (Instruction Set Architecture) 指令集架構

CPI in more detail

Performance Summary

• performance depends on
  • Algorithm: affects IC, possibly CPI
  • Programming language: affects IC, CPI
  • Compiler: affects IC, CPI
  • Instruction set architecture: affects IC, CPI, Tc

Tc => Clock Cycle 的週期時間

Power Trends

• Frequency => clock rate
• 當 clock rate 越高，功耗(power)就會越高，所以要降低電壓，才能使得功耗不要那麼高

Reducing Power

• Power Wall
  • 無法再降低電壓
  • 無法散熱

Solution of Power Wall

Uniprocessor => Multiprocessor

• Requires explicitly parallel programming
  • Compare with instruction level parallelism
    • Hardware executes multiple intstruction at once
    • Hidden from the programmer
  • Hard to do
    • Programming for performance
    • Load balancing
    • Optimizing communication and synchronization

SPEC CPU Benchmark

• SPEC 一個國際組織 訂定不同的 Benchmark 來評估自己的 Program
• SPEC CPU 2006
  • Summarize as geometric mean of performance ratios
    

SPEC Power Benchmark

• Power consumotion of server at different workload levels
  • Performance: ssj/sec
  • Power: Watts (Joules/sec)
    
    

Pitfall: Amdahl's Law

• Corollary: make the common case fast

Fallacy: Low Power at Idle

• Load 的百分比 等比下降，但是 功耗 不會等比下降
• 是還需要考慮、進步的

Pitfal: MIPS as a Performance Metric

• MIPS: Millions of Instructions Per Second
• 比較指令的數量，不能斷定誰比較好，還是要考慮執行時間