This paper considers the problem of scheduling hard real-time, periodic jobs on a multiprocessor while allowing imprecise computations. A highly dynamic job set is assumed, where limited a priori knowledge of a job set's behavior is available. The use of static partitioning schemes for such a job set is shown to lead to load imbalances and unnecessary error. Instead, a dynamic load balancing approach is used. A set of constraints are established so that balancing the load does not invalidate established schedules or increase the error. Finally a scheduling algorithm, based on these constraints is presented.