Workload characterization of uncacheable web content — and its implications for caching

Web caching and content distribution networks (CDNs) are two mainstream approaches to improve Web performance. Unfortunately, we have witnessed the inefficiency of these two approaches because of the rapid growth of uncacheable Web content, resulting from ever-increasing dynamic Web services and cache busting technologies. Although several approaches have been proposed for caching fragment-based dynamic Web content, such as edge side include and client-side include, we believe that a general understanding of uncacheable Web content characteristics is of great importance to future of content caching and delivery. In this paper, we characterized the uncacheable Web content by collecting an one-day HTTP traffic at a university setting. We found that (1) compared with cacheable content, uncacheable Web content now dominates the HTTP traffic (96.7%), with a larger average size (14K bytes of uncacheable vs. 5K bytes of cacheable); (2) although dynamic content generation consumes longer server processing time, network latency dominates the total user-perceived latency; (3) uncacheable Web content is distributed uniformly among multiple request/reponse pairs in one persistent connection; (4) personalized Web content is less than 10 % of total traffic, far lower than previous observations; (5) a considerable portion (42%) of Web content has their TTL values equal to zero. Among this content, we observed URL-alias contributes 38 % of total number of requests; (6) the university community has a time-specific access pattern during an one-day interval; and (7) P2P traffic is increasingly scattered among multiple ports (only 13 % from default ports for KaZaA traffic), which is a big challenge for P2P caching. Furthermore, based on these observations, we propose several future directions for efficient content caching and delivery of uncacheable content.