Abstract: A Web Cache cluster based on content-aware switching is implemented. It uses a layer-7 switch as the front-end node which transfers the client request to one of the back-end nodes, and its load is scheduled according to the content of HTTP header; some back-end node transfers the response to the client directly by the TCP handoff method, while the back-end need not recompile kernel because of ARP processing as is different from LVS. Tests, show high throughput performance, and the response time is reduced about 40% by the kernel-based Squid tool KSquid for non-cacheable object.