Re: LRU expiration age

On Tue, 14 Dec 1999, Henrik Nordstrom wrote:

[...]
> > OK, any suggestions as to the consensus for "reasonable" bounds on this
> > and any related measures?
>
> I would say a LRU age of about one week is quite well balanced for
> medium sized installations. More is probably not worth the effort.

Just to make sure I have this right:

This would mean that we are expiring from the disk cache objects which
have been "Least Recently Used" and that these have been sitting in the
cache unrequested for "about one week", to make room for new objects
(which have just been requested)?

And the stastical game we are playing here in the optimisation is
something like how long we need to keep unreferenced objects before the
probability that they will be referenced in the "near future" and still
"current" diminishes to a level "suitably" close to zero.

Lastly, the LRU age is not something we configure, but is the measure of
the age of objects dropped from the cache as above? The obvious
changeable inputs that influence the LRU age are the size of the disk
cache and the volume of (cacheable traffic) through the squid?

> This
> is assuming that your cache has filled up and then handled about the
> same amount of traffic it took to fill the cache, and that your Squid
> has been running some time with nominal load when you take the measure.

In my case, all these conditions should be met, with the observation that
the change of maximum_object_size from 4MB to 32MB is still within the LRU
age. Unsurprisingly, there appears to be a small upward creep in the
"Mean Object Size".

> This is a subjective opinion based on a number of things, and I have not
> measured it myself in real life.

Not to mention the dangers inherent in applying generalisations to
particular cases ;-)

Thanks,
Neale.
Received on Mon Dec 13 1999 - 23:41:28 MST