Parallelism

Disadvantages of Execution Policies

• May not have any effect

  • Not supported on some compilers
  • Not fully implemented on all compilers
  • May fall back to non policy version

• Extra overhead

  • The algorithm may start up new threads
  • The algorithm must manage these threads

When to Use an Execution Policy?

• Do not use an execution policy if

  • Your code has to be portable to other compilers
  • The task is essentially sequential
  • Operation order is important
  • The algorithm call throws exceptions
  • Unless immediate termination is acceptable
  • Preventing data races costs more than not using an execution policy

• Consider using an execution policy if

  • Measurement shows a worthwhile improvement in performance
  • It can be done safely and correctly

Which Execution Policy to Use?

• Sequenced execution is mainly used for debugging

  • Same as non-policy, but
  • Allows out of order execution
  • Terminates on exceptions

• Parallel unsequenced execution

  • Has the most potential to improve performance
  • Also has the strictest requirements. Use when
  • Data races cannot occur
  • Code does not modify shared state

• Parallel execution

  • Use when vectorization is not safe
  • Data races cannot occur, but code modifies shared state

• Unsequenced execution (C++ 20)

  • Can be used with single threading
  • If code does not modify shared state