The service overlay network (SON) is an effective means to deliver end-to-end QoS guaranteed applications on the current Internet. Duan et al. (2002) address the bandwidth provisioning problem on a SON, specifically, in determining the appropriate amount of bandwidth capacity to purchase from various autonomous systems so as to satisfy the QoS requirements of the SON's end users and at the same time maximize the total revenue of operating the overlay network. In this paper, we extend the concept of the service overlay network. Since traffic demands are time varying and there may be some unexpected events which can cause a traffic surge, these will significantly increase the probability of QoS violation and will reduce the profit margin of a SON. To overcome these problems, we propose to replicate services on the service gateways so as to dynamically adapt to these traffic surges. We show that the service replication problem, in general, is intractable. We propose an efficient service replication algorithm which replicates services for a subset of traffic flows. Under our replication strategy, one does not need to increase the bandwidth capacity of underlying links and at the same time, be able to increase the average profit for the overlay network. Experiments are carried out to illustrate that replication algorithm provides higher flexibility during traffic fluctuations and can quickly find a near-optimal solution.