In this paper, we study the problems of preemptive and non-preemptive online scheduling of jobs on unrelated machines in order to minimize the average time a job remains in the system. Both problems are known to be non-approximable by a constant factor. However, the preemptive variant has been extensively studied under the different resource augmentation models. On the other hand, the non-preemptive variant is much less explored. An O( 1/epsilon )-competitive algorithm has been presented in [7] for the non-preemptive average flow-time minimization problem on a set of unrelated machines if both an epsilon-speed augmentation is used and an epsilon-fraction of jobs is rejected. We are interested here in exploring the power of the rejection model and, mainly, in eliminating the need for speed augmentation in the latter result. On the road to this, we show how to replace speed augmentation with rejection in the preemptive variant. Our analysis is based on the dual-fitting paradigm.