Request Switching and Connection Multiplexing
NetScaler's fundamental performance secret is a patented traffic handling technique called Request Switching. It allows the NetScaler to decouple Layer 7 protocol requests from TCP connections. This allows for a more granular load balancing by making decisions for each individual Layer 7 request.
NetScaler combines Request Switching with Connection Multiplexing. Connection Multiplexing is a technique where warm connections are maintained with each of the Servers using Keep-Alives. The result is that server side connections are already scaled up to maximum speed/window size. The NetScaler then multiplexes requests from several client side connections and potentially several users to a single server using a single TCP connection.
If you consider the server side processing cost of setting up a TCP connection and tearing it down, the round trips needed to do so, not forgetting that each of these connections also has a memory cost, the benefits of request switching become immediately apparent.
Along with helping the server scale better, there are also other benefits to this technique. Because NetScaler is looking at traffic at the request level instead of the connection level, it is capable of offering better protection by looking at individual requests, instead of letting all traffic on a connection through and similarly, policies and load balancing decisions can be applied more granularly on a per request basis.
Note
The patent number for anyone interested in delving deeper is 6,411,986. This was later licensed to other vendors.
The following screenshot shows NetScaler Request Switching and Connection Multiplexing:
Tip
For troubleshooting, you might, on occasion, want to temporarily disable Connection Multiplexing, for example, to rule out issues of it not playing well with the Server. To do so, you can either use a HTTP profile with -conMultiplex DISABLED bound to the service, or by setting -maxreq to 1 on the service. Make this change at the service level to avoid impacting the performance of the environment as a whole.
