The TCP router nodes acts as Load Balancing front-ends to the Clustered web server. TCP request packets arrive at the public interface to the cluster (i.e. the Internet) where they are routed to an available processing node over the internal network connecting the router to the server nodes (for the SP, this is the CSS switch). The Router node executes highly optimized kernel-level code that allows a single Router to handle thousands of incoming connections per second. To further improve performance, the cluster is configured to allow reply packets to return directly to the external network.

For further scalability and availability, multiple TCP routers can be configured, and Round-Robin DNS can be used to distribute the load across the Routers. In this way, coarse-grain load-balancing can be achieved by the RR-DNS, while the Routers provide fine-grain load-balancing across the server nodes. The Routers can be configured to route to different sets of server nodes based on the target address or service requested. High Availability is achieved by automatic reconfiguration of routers / router take-over.

RR-DNS and TCP Router Configuration in Scalable Web Server

TCP routing has several advantages, including the following:

• Fine-grained Load Balancing : The cluster administrator has complete control over how the clusters resources are accessed from the outside. IP address caching by external clients cannot skew the load on the server nodes, as in the case of Round Robin DNS. The unpredictability of the Internet requires that the selection of processing nodes be handled at the server as connections are being established!
• Availability: In the event of a server node failure, the Router[s] are reconfigured to stop passing requests to the failed node. If a router node fails, RR-DNS is reconfigured to stop sending requests to the failed router and the router is also taken over on a spare/buddy, so that router failure is transparent to clients that may have cached the router address.,
• Security: The internal layout of the cluster is hidden from the outside world. Only the IP address of the Router[s] needs to be published.

The Router is based on code developed for the Encapsulated Cluster.

Rajat Mukherjee <rajatm@watson.ibm.com>

[ IBM Research home page ]

[ IBM home page | Order | Search | Contact IBM | Help | (C) | (TM) ]