Traditional multiprocessor real-time scheduling partitions a task set and applies uniprocessor scheduling on each processor. By allowing a task to resume on another processor than the task was preempted on, some task sets can be scheduled where the partitioned method fails. We address fixed-priority preemptive scheduling of periodically arriving tasks on m equally powerful processors. We compare the performance of the best algorithms of the partitioned and non-partitioned method, from two different aspects. First, an average-case comparison, using an idealized architecture, shows that, if a system has a small number of processors, then the non-partitioned method offers higher performance than the partitioned method. Second, an average-case comparison, using a realistic architecture, shows that, for several combinations of preemption and migration costs, the non-partitioned method offers higher performance.