ABSTRACT:: Requests for dynamic and personalized content increasingly dominate current-day Internet traffic; however, traditional caching architectures are not well-suited to cache such content. Several recently proposed techniques, which exploit reuse at the sub-document level, promise to add this shortcoming, but require a better understanding of the workloads seen on web sites that serve such content. In this paper, we study the characteristicsof a medium-sized personalized web site, NYUHOME, which is a customizable portal used by approximately 44,000 users from the New York University community. Our study leverages detailed server-side overheads, and the client-perceived request latencies. We then use these statistics to derive general implications for efficient caching and edge generation of dynamic content in the context of our ongoing CONCA project. Our study verifies both the need for and likely benefit from caching content at sub-document granularity, and points to additional opportunities for reducing client-perceived latency using prefetching, access prediction, and content transcoding.