Check out the new USENIX Web site.
... present.1
There are few exceptions to this rule such as responses to requests with AUTHORIZATION headers and responses with VARY headers that are not yet supported by Squid and are very rarely used.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
... age.2
Recall that the age of the copy is calculated as the maximum of the age and the difference between the current time and the DATE header value. Hence, if clocks at different hosts are synchronized, copies obtained from a cache should have positive age.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
... ages.3
Recall that when EXPIRES or MAX-AGE headers are not present, a heuristic is used to determine the freshness lifetime. The freshness lifetime is determined as a fraction of the difference between DATE and LAST-MODIFIED. In that case the TTL gets penalized twice for the copy's age: first, the freshness lifetime (fraction of the difference between DATE and LAST-MODIFIED subject to a limit) is smaller for older copies, and second, the age of the older copy is larger.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
... server.4
We distinguish a transparent reverse proxy cache from an origin server by issuing two consecutive GET requests within more that a few seconds apart to the same IP-address of the same server. A cache return a DATE header as obtained from the authoritative server. An authoritative server returns the current time according to its clock. Thus, we examined the DATE header value of the two HTTP responses. The same value indicated a cache. Two different values, where the difference approximates the time between our two HTTP GET request-response pairs, suggest origin server, very short TTL (few seconds or less), or a non-cachable object (that is, no age-penalty effect).
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
... proxy) 5
Interestingly, analysis shows that otherwise the age penalty is higher [3]
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
...  6
We conducted the following measurements with several URLs: We compared the latency of 500 GET requests of the URL from the same CDN server when (i) the same ARL was used each time (ii) a different ARL was used each time (e.g., by replacing the string e0a3b4215f9e4b). The cumulative latency in the latter measurement was 25%-40% slower than in the first. This suggests that when seeing a new Akamization, the CDN server obtains a new copy of the URL from the origin server prior to sending the response.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
... URL.7
As of 11/2000, Akamai seem to use HTTP redirect (302 response code) to the original URL in some cases when a new Akamization is seen.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
... server8
This is true in all cases except when the EXPIRES header is used in a relative way and there is no MAX-AGE directive (see Section 2), and when clocks on the origin site and the Akamai server are not synchronized
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
... (DIRECT)9
no requests were forwarded to a parent cache
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Edith Cohen
2001-01-16