That's often my thought too. Surely we can have the best of both worlds though? How far has virtualisation come in _simulating_ slower machines, whether it be cpu, network, memory or storage constraints?
Not good enough. Caches are the killer. A coworker of mine committed code that ran fine on his machine, but was rejected by our nightly tests on older machines. It turns out it wasn't that long ago that sub 1MB caches were standard, but his machine has 8MB of cache. This meant about a 20x reduction in performance on this algorithm (though some of that was also ALU performance).
