MPLS Based Web Switching

In this project, we are combining two emerging trends in the area of networking :
(a) the increasing use of MPLS (Multi-Protocol Label Switching) in core and access networks, and
(b) specialised layer4-7 web switches in front of server farms for high volume web-sites.

Web-switches serve four key functions with respect to routing user requests :
(a) content routing, where all web-content is partitioned amongst servers and a user request has to be routed to the specific server containing the requested content,
(b) load balancing user requests across servers within a cluster,
(c) directing user requests within a common session to the same server ("affinity") and
(d) routing user requests differently based oin the origin of the request ("service differentiation").

Typically, web switches either use TCP/IP headers (layer 4) or application layer header for dispatching user requests. Layer4 switches are suitable for load-balancing and high performance but are unable to support content routing. Content routing requires terminating incoming TCP connections and examining higher-layer headers. As web application requirements evolve, there will be a need for more sophisticated dispatching since the approach of layer4 switching of examining only TCP/IP headers provides insufficient functionality, while layer7 dispatchers suffer from scalability and performance limitations due to their requirement of terminating incoming connections. In this project, we are working on a approach that, when combined with an intelligent client-side proxy, can implement a dispatcher using off-the-shelf MPLS switching hardware, while also providing the flexibility of a content-aware router.

MPLS has been proposed both as a means of speeding up packet processing and for establishing traffic-engineered paths within a network. Rather than use labels to express routing and forwarding policies, the key idea in this project is to map application layer information to enable high-performance and yet flexible, web-switching. MPLS provides for label stacking and we exploit this feature of MPLS to stack the label used for web-switching below the routing labels. Consequently, the intermediate MPLS network between a client-side proxy and the web-switch in front of a server farm, is unaffected by our use of MPLS labels for web-switching at the edge of the network. A mapping of labels to application-layer semantics is communicated by the dispatcher to the proxy through a control connection and this mapping can implement content-based routing, client-server affinity, client-specific service differentiation and server load balancing without requiring termination of TCP connections at the network dispatcher.

The main advantage of this approach is that it removes a critical bottleneck from the system, namely the need for terminating incoming TCP connections at a single point for dispatching user requests. Further, it lends itself to realization in a standard off-the-shelf MPLS switch, thus obviating the need for specialized, layer-7 web switching hardware.