Character of Scientific Computation

• Algorithms evolve, sometimes with physics.
  • Often not "canned" algorithms

• Large blocks of floating-point arithmetic.

• Testing on a new platform may be highly non-trivial.
  • Different machine arithmetic, non-identical answers

  • Physics codes can run happily with bugs
    • May require several long, expensive physics runs to notice the bug; more to figure it out.
      • Debuggers don't allow saving state

    • Small test-case runs may be useless for debugging
      • Small enough for 1 MPP node may be physically absurd.

Generated with CERN WebMaker