How to do (parallel) performance analysis

• The first step is to verify that your code is correct
  • construct a test that convinces you.
  • Eg, if you compute something known by other means (like "pi") compare with it
  • print the result of the test with enough comments that you will remember what it was a year later
• The second step is to time your (sequential) code
  • pick a timer with sufficient resolution
  • decide which segments of the code you want to time separately
  • repeat short segments, or choose a large enough data size, to allow your timer to resolve them
• The final step is to time your parallel code and do a performance prediction graph
  • A trick of the trade: synchronize all processors before you begin timing.
  • use log-log plots
  • separate timeing-data gathering and printing from the timing sections
• Helpful hints
  • label all your graphs (both axes, with units, and a title) so you will remember what it was a year later
  • make your "output" a complete report (give the name of the timer used)
• Communication
  • At least initially, you should always submit the complete code listing
  • At least initially, you should always print out timing (and other data) from all processors to verify what is going on