Pitfalls I

Pitfalls I

• Replacing a good sequential algorithm with a terrible parallel one
  • Classic example is using Jacobi iteration instead of hierarchically-preconditioned conjugate gradient
• Assuming floating point associative
  • common mistake for all-reductions
  • Lack of bitwise-identical results
    • For example, don’t assume that each process will compute exactly the same values given the same data unless you can guarantee the same order and intermediate forms (beware compilers!)

Previous slide

Next slide

Back to the first slide

View Graphic Version