Computer abstraction and Technology

defining performance can be tricky!

  • Response Time : How long it takes to do a task
  • Throughput : Total work done per unit time
    • more emphasized in parallel programming
  • replacing processor with a faster version
  • add more processors -> more tasks in the same time (increasing throughput)

Relative performance

Performacne = 1 / (Execution time)

if X is $n$ times faster than Y, the excution time of Y takes $n$ times longer

Measuring execution time

  • Elapsed time (Total response time)

    • includes CPU time, disk access, network access, os overhead, idle time, etc.
    • measuring the whoe system performance at once

  • CPU time (time spent processing a given job)

    • user CPU time + system CPU time
    (Clock Rate) = (Clock Cycle) / (CPU time)
(CPU time) = (Clock Cycle) / (Clock Rate)

    (CPU clock cycles used) * (clock time per cycle) = (cpu clock cycles) / (clock rate)

    • Improving performance

      • reducing clock cycle #
      • increase clock rate
      • clock rate vs cycle count? : increagin clock rate can cause more clock cycles

    • Instruction counts: Determined by program, ISA, compiler, etc.

    (Clock cycles) = (Cycles per Instruction) * (Instruction #)
(CPU Time) = (Instruction #) * (CPI) / (Clock Rate)j

    Different instructions have different CPI: average CPI utilized

    • $(\text{average CPI}) = \displaystyle \sum{(\text{CPI}_{\text{instruction}})} \cdot (\text{Relative frequency of instruction})$

CPU Clock

digital hardward generating constant-rate clock

  • synchronized circuits utilizing the clock cycles of cpu

Performance summary

$$ \text{CPU time} = \frac{\text{instruction #}}{\text{Program}} \cdot \frac{\text{Clock cycle #}}{\text{Instruction #}} \cdot \frac{\text{clock cycle time}}{\text{Clock cycle #}} $$

  • Power consumption
9Power consumption) = (capacitive load) * (frequency) * (voltage) ** 2
  • power wall! (around 2003)
    • we cannot reduce voltage: reliability problems
    • we cannot reduce heat: temperature of cpu increase as as power consumption increases

introduction to multicore processor (from unicore-processor)

  • Multiprocessors: hard to program (explicit parallel programming)
    • c.f. instruction level parallelism (ILP)

Amdahl's Law

$$ T_{\text{improved}} = \frac{T_{\text{affected}}}{\text{improvement factor}} + T_{\text{unaffected}} $$

we should make the affected portion large enough to improve the overall performance!