Our approach to dealing with I/O is somewhat unsatisfactory, as we do not consider the I/O time a task requires until after it actually occurs. A better approach would be to model the probability distribution of task I/O requirements for each task type and use this distribution to compute a more optimal schedule at the outset of the task. This requires a more complicated model of speed and voltage scheduling, and consequently a more complicated solution to computing an optimal schedule than PACE currently uses.