A method and apparatus providing highly scalable server load balancing are
disclosed. Data packets from a client are routed through one or more
routers to a server load balancer, which is selected from among a
plurality of server load balancers in a network. In response to receiving
a request packet, a particular server site to process the client request
is selected. A first path to a second router associated with the
particular server site, and a second path to a server load-balancing
device associated with the second router, are determined. A mapping of
flow identifying information, associated with the packet, to a first
label value that identifies the first path and to a second label value
that identifies the second path, is created. The first label value and
the second label value are stored in the packet. All subsequent packets
associated with the client request are forwarded to the server
load-balancing device based on looking up the first label value and
second label value in the mapping. As a result, a network is scalable to
process and load-balance numerous client requests, which are efficiently
routed to the site, server load-balancer, and server that are handling
the request.