We investigate particularly simple algorithms for optimizing the tradeoff
between load imbalance and assignment overheads in dynamic multiprocessor
scheduling scenarios, when the information that is available about the
processing time of a task before it is completed is vague.
We describe a simple and elegant generic algorithm that, in a very
general model, always comes surprisingly close to the theoretical optimum,
and the performance of which we can analyze exactly with respect to constant
factors. In contrast, we prove that algorithms that assign tasks
in equal-sized portions
perform far from optimal in general. In fact, we give evidence that the
performance of our generic scheme cannot be improved by any constant factor
without sacrificing the simplicity of the algorithm. We also give lower
bounds on the performance of the various decreasing-size heuristics that
have typically been used so far in concrete applications.