I have studied the effect of other factors that influence scaling of cluster dynamic content web sites: admission control, load balancing and transparent query caching.
I have shown that the choice of load balancing and in particular Locality-Aware Request Distribution Scheme (LARD), which has been shown to be successful for load balancing static content requests in a cluster have little impact in a dynamic content cluster. The explanation arises from the basic differences between static and dynamic content. The latter is typically more CPU intensive, and has a lot of locality. In on-line shopping, bestsellers, promotional items and new products are accessed with high frequency. Similarly, the stories of the day and the items on auction are hot objects in bulletin board and on-line bidding applications, respectively. Thus, a lot of complex queries access hot rows in a few tables which become replicated in most memories independent of policy. Secondly, given that traffic is non-uniform due to client think times, limiting database congestion during bursts of traffic through admission control becomes more important than the particular choice of load balancing.
Lastly, I have showed that query result caching and clustering through replication complement each other well, and can be used in combination to improve performance. The benefits of query result caching are limited for workloads with high rates of content modification due to frequent cache invalidations. In these cases, conflict-aware replication is beneficial. On the other hand, for a given cluster size, caching can significantly improve performance over clustering under workloads with few modifications.