Presented is a system and a method for load balancing multiple globally-dispersed servers based on client-centric performance criteria. The infrastructure of the system includes load balancing domain name servers (DNS-LBs) deployed in close physical proximity to the Internet service providers' points of presence. The DNS-LBs are then able to monitor the performance of the servers from a location close to the clients, which allows the DNS-LBs to select a server that will yield the best performance from that location for the client. A second level of the infrastructure utilizes domain name servers (DNS-Bs) that are deployed on the Internet backbones and regional provides. The authoritative domain name servers (DNS-As) for the servers to be load balanced refer all name queries to these DNS-Bs. The DNS-Bs then refer the queries to one of the DNS-LBs based on a mapping of the DNS-ISP address to its physically proximate DNS-LB. The DNS-LB then returns the IP address of the server that will provide the best performance from that location.