Processor Performance

Attributes

• Execution/Response Time
  • The time between the start and completion of a task
  • Improving response time improves throughput
• Throughput
  • Total amount of work done in a given time

Comparing Performance

• $P=\frac{1}{E}$ (P: performance, E: execution time)
• “X is n times as fast as Y” → $\frac{P_x}{P_y}=\frac{E_y}{E_x}=n$

Measuring Performance

• Elapsed time
  • Total time to complete a task
  • Includes disk accesses, memory accesses, I/O, OS overhead, etc.
• CPU (Execution) Time
  • The time the CPU spends computing the task
  • Does not include time spent waiting for I/O or running other programs
• User CPU Time
  • CPU time spent within the program itself
• System CPU Time
  • CPU time spent by the OS performing tasks on behalf of the program
  • Lines of this is pretty blurry

Amdahl’s Law

• When you’re given the time something is taken as a part of a whole and then how much that part is sped up by
• Ex: 60 seconds of 100 seconds is improved by 3x

$$T_{\text{improved}}=\frac{T_{\text{affected}}}{\text{improvement factor}}+T_{\text{unaffected}}$$ $$\frac{\text{Performance with}}{\text{Performance without}}=\frac{1}{(1-\text{fraction affected})+\frac{\text{fraction affected}}{\text{improvement factor}}}$$

Application: If you’re going to improve something, improve the thing that you use the most

Can deal with multiple performance changes by adding more of nested fractions (the 1 - fraction will be 1 - the sum of all the affected fractions)

Impacts

$\text{Time}=\frac{\text{Instructions}}{\text{Program}}\times \frac{\text{Clock Cycles}}{\text{Instruction}}\times \frac{\text{Seconds}}{\text{Clock Cycle}}$

Clock

• Clocks synchronize the hardware
• Clock Cycle
  • Time intervals in which the CPU can do an action
  • Synchronizes all the hardware
• Clock Rate
  • Clock cycles per second
  • Measured in (giga)hertz
  • Inverse of clock cycle time
• $\text{CPU Time}=\text{Clock Cycles}\times \text{Clock Time}=\frac{\text{Clock Cycles}}{\text{Clock Rate}}$
  • This is one way that faster hardware can make a program run faster
• Increasing clock rate
  • Can improve performance in exchange for using more Power
  • Can only go so fast because eventually you will hit the limits of the hardware (adders can only add so fast)

Instruction Performance

• Different instructions take different a different number of cycles
• Different factors can impact Instruction Count (IC)
• Clock Cycles per Instruction (CPI) = Average number of clock cycles an instruction takes to execute for a program of program fragment
• $\text{CPU Time}=\frac{\text{Instruction Count}\times \text{CPI}}{\text{Clock Rate}}$

Factors

• Algorithm: IC, possibly CPI
• Programming language: IC, CPI
• Compiler: IC, CPI
• Instruction set architecture: IC, CPI
• Chip implementation: CPI

Reporting Performance

• There are chip independent test suites
• Take the weighted arithmetic mean of the execution times
  • weighted by the ratio of how frequently the tasks occur relative to each other
• You can use the geometric mean to compare performance to a reference machine

Millions of Instructions Per Second (MIPS)

• Really bad at comparing processors
• Doesn’t account for
  • Difference in Instruction Set Architecture between computers
  • Differences in complexity between instruction
  • Differences in CPI between programs