... cache¹

Our study in [35] indicates that a first HTTP request to a server in a time period is more likely to take longer than subsequent ones.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... policy.²

We specify optimal offline algorithm in [25].

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... URLs³

The original logs had about 5 million different URLs, most of them requested only once. However, most of the requests are for frequently requested URLs, so reducing the miss rate of those is the goal.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... PASSIVE.⁴

We remark that these policies are able to achieve ``continuous'' tradeoffs by mixing two consecutive integral values of the respective parameter, for example, applying RECENCY(1) on some URLs and RECENCY(2) on others.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... framework.⁵

Recall that under the refreshment framework, objects must be kept fresh continuously till the following ``hit.'' The optimal unrestricted policy, which validates objects just before they are requested, incurs no overhead.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... renewals⁶

It could be less than k since we did not perform renewals passed the termination time of the log.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.