1.6 Performance

• In my opinion this is the important section of this chapter.
• We will have played with some of these ideas in the math section.
• How can we measure performance?
  • Clock speed?
    • Within a processor family somewhat
    • Across processor families? NO
      • A hummingbird has a heart rate of about 1000 BMP
      • An elephant has a heart rate of about 30 BMP
      • Which do you want to haul wood in a jungle?
      • I got these from here.
  • How about instructions per second?
    • $ \text{IPS} = \frac{\text{cycles}}{\text{second}} \times \frac{\text{instructions}}{\text{cycle}}$
    • But it is most likely these two factors are directly related.
    • Ie the difficulty of the instruction determines the length of the clock cycle.
    • RISC chips do lots of little instructions very quickly (Hummingbird)
    • CISC chips do a few big instructions more slowly (Elephant)
    • So this is just another measurement of processor speed.
  • Specialized measurements ie a test suit
    • Companies have been known to cheat.
• They give an interesting table in Figure 1.14, page 29
  • Comparing the throughput of various airplanes.
    • The Concorde was great for moving a few people very quickly
    • The airbus is great for moving many people a long way.
• Time
  $\text{seconds} = \frac{\text{seconds}}{\text{cycle}} \times \frac{\text{cycles}}{\text{instruction}} \times \frac{\text{instructions}}{\text{program}}$
• A note about compilers
  • These generate code in machine language
  • Different compilers will generate different code.
  • A compiler can optimize code.
    • This rearranges, substitutes and performs other tricks on the code.
    • The code still produces the same results but will run faster.
    • Some compilers are better at optimization than others.
    • I don't know about the validity, but here is a Comparison of some compilers.
  • So compilers can change the instruction mix, hence the CPI
  • They can also change the instruction count
• The ISA
  • This determines CPI and Clock Speed.
  • But also the instructions available for the compiler.
• The programming language
  • May support abstractions that require more instructions.
  • IE directly contribute to the instruction count.
• The programmer
  • Instruction count, instruction mix.
• The algorithm
• The only real measurement of performance is the time for your application
  • With the provided compiler
  • On the machine you want to test.