Performance

• How can we effectively measure performance?
  • What is performance?
  • Look at Figure 1.13 page 28
    • We assume we are talking about transporting passengers.
    • But what different criteria can we have for transporting passengers?
      • Get 100 people there as quickly as possible.
      • Get 100 people there as far as possible.
      • Get 450 People there.
    • Notice that for each of these tasks, a different plane would be selected.
    • The parallel we can draw here is that many people have different problems that require different hardware solutions.
  • We will define performance as the inverse of execution time.
    • Performance_x = 1/Execution Time_x
    • As execution time (t) increases, 1/t decreases
  • If P_x > P_y then T_x < T_y
  • (or) T_Y/T_X = n
  • If X is n times faster than Y, P_X/P_Y = n
• Modeling execution time
  • Execution time (T)
  • Instructions Executed = I
  • Cycles Per Instruction = CPI
  • Clock Speed = S (cycles per second)
  • T = I * CPI * 1/S
• How can we effect execution time?
  • Who controls I?
  • Who controls CPI?
  • Who controls S
  • Look at page 38
• CPI and Instruction mix
  • Assume machine A has a clock rate of 1.4GHz.
  • Assume the program has 1x10¹² instructions
  • Assume the following

    Instruction Class	CPI	% used
    A	3.1	30%
    B	1.0	50%
    C	4.0	20%

  • Find the CPI.
  • Find the execution time.
  • What if the instruction mix is changed to 25%, 60%, 15%
  • What if the CPI for class B is raised to 1.1, but the other two are reduced by .5?
  • What happens if the clock speed is raised to 1.5GHz?
  • What if the instruction count is reduced to 5x10¹¹
  • How can each of these be accomplished and where?