RE: Storing partial responses

From: Robert Collins <robert.collins@dont-contact.us>
Date: Wed, 13 Dec 2000 16:36:23 +1100

> -----Original Message-----
> From: Alex Rousskov [mailto:rousskov@measurement-factory.com]
> Sent: Wednesday, 13 December 2000 4:20 PM
> To: Robert Collins
> Cc: squid-dev@squid-cache.org
> Subject: RE: Storing partial responses
>
>
> On Wed, 13 Dec 2000, Robert Collins wrote:
>
> > Hmm. Well I am definately heading to a (2) approach, because
> > performance tuning will need both FS and meta level activities.
>
<snip>
>
> An orthogonal example would be the delegation of
> responsibilities between hard drive controllers (fast operations on
> semantic-less easy-to-handle blobs/blocks) and Unix file systems
> (relatively slow operation on files). One could argue that if disks
> knew more about files, they would perform "better".

So the question I have: is the existing squid FS code a low level 'tool'
like the mechanicisms of a hard disk, or a medium level interface like
the unix file systems? I suspect it is at the file system level.

>
> In other words, sometimes it is better to keep the lower
> levels of the design fast and simple, while making slow but smart
> decisions at the meta-level. You do loose efficiency because of the
> communication between the layers, but you gain in overall design
> flexibility, which often proves more important, even performance wise.

I agree. See for example the acl result caching I've done in
auth_rewrite. It wasn't obvious that that could be done until some
sensible layers had been created, (namely,
access<->authentication<->user directories).

In the squid case here is how I see it layering out (corrections
requested :])

meta level - rfc 2616 decisions/acl/performance choices for request
satisfaction/choice of different FS's based on the FS capabilities.

---
fs level - efficient and fast object storage. Often tuned for a specific
purpose - ie small objects/big objects. The tuning implies smart choices
about object storage locations.
---
o/s level - high speed physical disk access.
Finally I think that the meta level will get too bogged down in the
mechanics of partial response storage if it is not abstracted (as per
your (2) before).
Rob
Received on Tue Dec 12 2000 - 22:40:10 MST

This archive was generated by hypermail pre-2.1.9 : Tue Dec 09 2003 - 16:13:03 MST