Re: Web caching research

Trevor Strohman wrote:

> I'm writing this message to you because you seem to be on the
> forefront of Squid development. I've been wondering recently
> about web caching algorithms and specifically why invalidation
> protocols aren't widely used.

Cache replacement:

LRU is choosen since
1. It is very simple
2. The difference between the cache replacement polices is only seen on
smaller caches. On a larger cache all perform close to identical.

Freshness policy:

This is a complex issue. There are three goals here
1. The end user should preferably never see stale documents
2. Quick response times
3. Little network overhead

and another important factor is code complexity.

Adding yet another protocol makes things much more complex, and it is
questionable if such a protocol can be found that justifies the amount
of network overhead that is needed compared to the freshness level seen
by the end user.

> Could you point me to any information on why Squid is implemented
> entirely based on time (TTL, Expires)?

No one has bothered with looking into another implementation. And
another issue here is the amount of metatada that we needs to keep about
a object.

We will probably add support for the cache new control directives from
HTTP 1.1 at some day that gives the origin server more control of the
freshness level and/or hit metering.

> Also, it has occurred to me that for certain types of often-cached
> documents (like a news page) that caches could update themselves by
> sending differences between the old version of the page and the new
> version, instead of a whole new copy. Is this feasible?

It depends on the situation. For a general cache no, but if this kind of
documents is a large percentage of the cache then it may be feasible.
There are however two major issues:

1. There is no defined standard for sending updates.

2. It adds a level of complexity to the cache. Special care must be
taken to ensure that the document is properly updated.

There are probably larger gains seen from adding compression rather than
delta updates.

---
Henrik Nordstrom
Spare time Squid hacker