Web Caching and Replication

6.4. Object Popularity

The Web exhibits very uneven object popularity: most accesses are for a relatively small set of objects, while the vast majority of objects are hardly ever accessed. Arlitt and Williamson [1996] examined six different Web sites and found that 10 percent of objects accounted for 80 to 95 percent of all requests received by each site. More specifically, several studies [Glassman 1994; Almeida et al. 1996; Cunha et al. 1996; Barford et al. 1999] observed that object popularity conforms to Zipf-like distribution [Zipf 1949]. Applied to the Web, Zipf-like distribution states that the popularity (that is, the frequency of accesses) of the i th most popular object is proportional to 1/i^α, for some constant α between 0 and 1. In ...