Optimizing TCP Forwarding

The continued growth of the web places ever increasing performance demands on web site front-end appliances. In many cases, these appliances have to forward network traffic to and from web servers at transport and application levels utilizing complete TCP/IP stack processing, which could easily make the front-end appliance a bottleneck for a web site. This paper describes four novel optimizations of the TCP/IP stack processing for a TCP forwarding appliance: acknowledgement aggregation, fast path for incoming packets, double allocation avoidance in TCP module, and packet reuse. These optimizations are applicable for different throughputs and MTU sizes on the forwarding path when traditional approaches for performance improvements, such as TCP splicing, could not be applied. These optimizations are implemented in context of a web booster appliance that modifies web site traffic to decrease the cost of the network processing on a web server. These four optimizations, applied together with device driver polling, result in four-fold improvement in appliance throughput compared to a base case.