Performance

• This is 1.6 in the book.
• What does performance mean?
• They give a nice chart in the book

  • Airplane	Capacity (passengers)	Range (miles)	Speed (MPH)
    Boeing 777	375	4630	610
    Boeing 747	470	4150	610
    Concorde	132	4000	1350
    Douglas DC-8-50	146	8720	544

  • Why the differences?
  • Are all planes equally good?
  • What is best for a "bus in the sky"?
  • How about getting there fast?
  • What about something for a small intercontinental sales team?
  • What other measures might be important?
  • They give one other number

  • Airplane	Throughput
    Boeing 777	228,750
    Boeing 747	228,700
    Concorde	178,200
    Douglas DC-8-50	79,424

  • Where throughput is passengers x mph
  • Is there a problem with this measurement?
• When attempting to define performance in a CPU (or computer in general) there are a number of problems
  • Is clock speed a good measurement of performance?
  • How about instructions per second? (MIPS was once a standard)
  • Clock Rates, CPI and other measures are not valid comparisons across different instruction sets. Only in the same family.
  • Section 1.9 discusses a current benchmark set for CPU performance
  • Fallacy and Pitfalls discusses how MIPS is a bad measurement.
  • Since this is related to marketing, sometimes companies have behaved poorly.
• In the end, the speed at which your program is executed is the best valid measurement of performance.
  • Simple or complex instruction set
  • Clock speed
  • And other considerations
• Execution Time = seconds/cycle x cycles/instruction x instructions/program
• Performance = 1/Execution Time
• If machine X is n times faster than machine Y
  • PerformanceX/Performance_Y = n
  • Or Execution_Y/Execution_X = n
• Problem 1.5 page 55

          Clock(GHz)    CPI
  P1:      3.0          1.5
  P2:      2.5          1.0
  P3:      4.0          2.2
  
  I have:
        cycles/instruction  (CPI)
        cycles/second       (Hz)
  
  I want:
        instructions/second (Performance)
  
        A cycle    instructions      A instructions
        ------- x  ------------  =   --------------
        second      B  cycle        B   seconds
  
  P1: 
        3x109
        -----  = 2x109
         1.5
  
  P2: 
        2.5x109
        ------  = 2.5x109
         1.0
  
  P3:
        4.0x109
        ------  = 1.8x109
         2.2
  
  Relative performance of P1 to P2 is
  
           PerformanceP2    2.5x109
  	 ------------- =  ------- = 1.25 
           PerformanceP1    2.0x109
  
  	Or P2 is 1.25 times as fast as P1.
  	Or P1 is 1.25 times slower than P2.
  	 

• Example, Question 1.12 page 58

            Clock    CPI    Instructions
        P1:  4GHz    0.9     5.0E9
        P2:  3GHz    0.75    1.0E9
  
        Which has better performance and why?
  

• Example: One I made up

      Assume multicycle  CPU  has the following  instruction mix for a given program.
            Instruction Class         Cycles       Frequency
  	       A                      1             20%
  	       B                      2             40%
  	       C                      3             30%
  	       D                      4             10%
  
      Find the average CPI
  

• They have a discussion on page 38 that is important
  • What affects the speed of a program?
    • The ISA (Instruction Set Architecture)
      • Clock Speed
      • CPI
      • Instructions
    • Compiler
      • Choice of instructions -> instruction mix -> CPI
      • Instruction count
      • Memory access -> CPI
    • Programming Language
      • Instruction Count
      • Choice of instructions
      • Abstraction -> more function calls
        • Think of a OO program vs a structure program.
    • Algorithm/Programmer
      • Instruction count (BIG TIME)
      • CPI (instruction choice)
        • x = x/2;
        • x = x >> 1