On modern cpu architectures it can be very difficult to predict running times, due to instruction pipelines and caches. So I don't think this approach will work unless a lot more complexity is added.