Calculating Performance

Notes

• Hopefully last time we came up with a list of items that might be considered for measuring performance in a computer
  • CPU speed
  • Memory Capacity
  • Bus Bandwidth
  • Capacity for peripherals
  • ?
• H & P focus on program execution time
  • Is this strictly related to clock speed?
  • No, clock speed is just one component.
• Execution Time Components
  • CPU speed or cycles per second
  • Number of instructions executed
  • CPI or cycles per instruction.
• Clock speed:
  • Measured in Hz.
  • Determined and fixed
  • Not much you can do about this.
• Instructions
  • Measured in instructions
  • Based on the problem/program/algorithm/compiler
  • You have some control over this
• CPI:
  • Determined by the operations you perform.
  • Based on the problem/program/algorithm/compiler
  • You have some control over this
• perf
  • Look at the man page (man perf)
  • This is more accurate on intel than arm
  • perf stat -e cycles,instructions prog
• An experiment with perf
  • I have written several example programs.

  • Program	Operation	Instructions	Cycles	CPI
    intAdd	a = b + c	9,601,972	9,679,109	1.01
    intMult	a = b * c	8,601,971	8,771,291	1.02
    intDiv	a = b / c	9,601,971	16,167,373	1.68
    floatAdd	a = b + c	9,601,971	10,150,683	1.06
    floatDiv	a = b / c	9,601,976	26,193,372	2.73
    arrayAdd	a[i] = b[i] + 7	1,048,263,321	1,257,159,126	1.20
    arrayDiv	a[i] = b[i] / 7	1,055,263,333	1,253,667,290	1.19

  • These are not perfect, they are just to show
    • Int add and multiplication are fairly similar
    • Int division is yucky
    • Array access is somewhat in-between , and dominates the operation
    • Floating point add, and multiply are not too bad
    • Floating point division is very yucky
• In the end, my experiment is flawed because
  • I have not really studied perf, so I might not be using it correctly
  • I have not automated data collection and processing, so there could be hand copy and paste errors
  • I am allowing the compiler to move to assembly language, so I could be encountering some unexpected optimization
    • Definitely do not us -On when compiling.
  • Cache could have an impact here, I was definitely trying for that in the array programs.
• But it should be clear that different instructions have different CPI's
  • And that this is somewhat controllable by the programmer.
• Note, that the cpi does not have to be greater than 1
  • This seems strange, but we might be able to execute more than one instruction at a time!
  • See me at the end of the semester.
• I will get accounts set up on Archimedes soon, you should
  • Consider an hour or so of experimentation to see if Intel and ARM are similar
    • Note: The ARM family handles instruction tracing much different than the Intel family
    • Do to this, the results reported by perf on an ARM machine are very different.
  • Try this on another Intel or AMD architecture
    • Mirkwood has Intel(R) Core(TM) i7-2600 CPU @ 3.40GHz
    • cat /proc/cpuinfo | grep "model name" on linux
  • Experiment with other programs/instructions. See if you can find a combination that is worse than floatDiv.cpp