We also discussed some cache policies. To model an existing TiVo user, we came up with a scheme where the user only has awareness of a certain percentage of the object space. Anything outside this portion of the space will be ignored. Anything inside this portion of the space will be placed into the cache, removing those objects of lessor utility. This models the current TiVo replacement policy which requires the user, who only knows about a certain number of shows, to delete cached shows to make room for a show the user would like to watch. In contrast, a simple automatic cache policy would know about the entire object space, and remove objects of lessor utility to make room in the cache, but would have some margin of error for determining object utility. Comparing the average utility curves as margin of error ranges to the average utility curve as viewer awareness ranges will provide an interesting view of when the automatic cache policy becomes useful. A natural extension of the automatic cache policy is one where objects are compressed to try and maximize the number of objects and overall utility.

Unfortunately, neither of us has any idea of how to even start looking at what sort of distribution, consumption, or utility curves would accurately model real television viewing behavior. The object arrival rate is relatively constant since there are a set number of channels broadcasting at any given time. We can sort of guess as to what sort of curves are appropriate for the other things, but there's nothing we can point at to support our choice of curves. Mark Lindsey had looked into this a little bit, so Ketan suggested I email him to see if he had any idea.

Post a comment