Web Caching

Hit Ratios

In this section, I use the Anon-U trace files in two different simulations. We’ll see how the number of cache clients affects hit ratio and how the hit ratio changes as a function of the cache size.

For hit ratio versus number of clients, I used one day’s access log from the Anon-U data. I am restricted to a single day because the client IP addresses are not sanitized consistently between log files. The access log I used has 1.7 million requests from 2,415 different clients. I can’t tell from this data which of the objects are uncachable, so in the simulation all URLs are uniformly uncachable with a 25% probability.

Given a trace with 2,415 clients, how can we calculate the hit ratio for one, two, or any particular number of clients? My approach is to first create a random ordering of clients. Then I run a program that reports the hit ratio for requests from the first client, the first two clients, the first three clients, and so on, all the way up to the maximum. For small numbers of clients, the hit ratio varies greatly depending on the ordering of clients. In one case, the first 20 clients might have a 20% hit ratio, while another results in 90%. To improve the precision, I ran the simulation with 100 different client orderings and averaged the results. Note that, in these simulations, the cache size is infinite, and there are no If-modified-since or no-cache requests. The random seed is always the same so that the set of uncachable URLs remains consistent between ...