The replication of digital content over the Internet makes the identification of the best server an interesting problem. In this paper, our aim is to reduce the content transfer time from a server to a client with avoiding network and server congestion using an efficient, scalable server selection scheme. First, we review the limitations of the anycasting schemes proposed in the literature. Then, we propose an application-layer anycasting scheme which is a scalable, transparent, and dynamic solution. Our anycasting scheme is based on a metric which can predict the transfer time of a content transmitted using TCP. Our metric considers the server load, the server's maximum sending window size, and the client's maximum receiving window size. Also, it considers the critical performance parameters on the path server - client (i.e., the available bandwidth, the round trip time, and the packet loss rate). Our empirical results show the weakness of the other prediction schemes. Moreover, the results show that our proposed function predicts the transfer time of a content with an 8.13% accuracy compared to the real transfer time.