Re: Alternative Squid Filesystem

Adrian Chadd wrote:

> The difference here is that the object store being proposed isn't an
> exact LRU, its an average LRU. I actually like the idea. After some more
> preparation work is done with the current codebase so it will benefit
> from this kind of filesystem.

Different from the COSS implementation yes. The idea was however toyed
around with in principle.

Ideas toyed around with at that time:

* Plain FIFO

* FIFO with LRU extensions by copying referenced objects to the head

* Multilevel FIFO with LRU extensions, where referenced objects were
copied to another FIFO instead of the one where fresh objects are
stored. Applied recursively as many levels as seem fit. (Note: This can
approximate frequency based policies)

* Chunked filesystems where space is reclaimed in larger chunks based on
policy weights for the chunks.

* Chunked filesystems with repacking of objects

* Chunked filesystems and what to do with objects not fitting within the
selected chunk (i.e. semilarge object of unknown size)

And a couple of combinations of the above.

State logging was also covered to some extent. My idea was to instead of
keeping a large index keep a transaction log for the store close to the
data it refers to (i.e. in each chunk for the objects contained there).

How much of this I managed to communicate in the squid-dev discussions I
don't remember, but I think most of it was covered one way or another.
My archive of the discussion is a bit scattered and I might have lost
some parts (it was assembled many months afterwards from my mailboxes).
Unfortunately the squid-dev archives does not cover that time
(january/february 1999).

/Henrik
Received on Tue May 30 2000 - 18:16:12 MDT