digest-mesh, euro-mesh or whatever ... (fwd)

-- refresh pattern tricks and some data pointers might be of use to
   squid-users...

Alex.

---------- Forwarded message ----------
Date: Tue, 9 Feb 1999 12:03:29 +0100
From: "Javier Puche. CSIC RedIRIS" <javier.puche@REDIRIS.ES>
To: TF-CACHE@TERENA.NL
Subject: digest-mesh, euro-mesh or whatever ...

Hi all,

 I am very happy about the good will for peering with cache-digest
no-query (no ICP) that we talked about last meeting; so RedIRIS,
Surfnet, Arnes, Hungary and probably Nordunet (someone spoke for you
Ingrid :-) showed interest for starting it up. I expect more people to
join (janet, switch, dfn,...)

The only problems up to date are:

- With Squid 2.1 we still have the false-hit problem, the only
solutions are either wait to be running and have confidence on squid
2.2 or start right now with the delicate miss_access and trusting we
are not going to abuse other caches. Opinions ??

- Once the false-hit problem dissappears it is not that dramatic to
have the refresh patterns in sync, but it stills seems appropiate to do
it. With digests only the default refresh pattern is applied (thanks
again to Martin for the documentation), mine ones are '0 40% 1440'
but I would move to whatever is agreed here, I would suggest however to
move to something a bit more relaxed. (it is sad with the 0 to see in
the logs repeated access to something that was fetched 1 second ago,
and the users know about the shift-reload anyhow)

- The experience we are having here with digest after a survey of 10
caches is that many regional caches get really poor quality digests
(over 90% rejected, 8 bit_seq, ...). We are running experiments to see
how to get better quality digests (and read Martin doc also), trying to
correlate refresh patterns and other config params with the util%,
reject% and bit_seq. We are getting the best results when the 'max'
part of the refresh patterns gets close to the LRU expiration age and
the min is not very low. Following are the data collected by one
regional cache (obtained by breton@cica.es), it's not enough to draw
conclusions but just to show what I am talking about:

refresh . 1440 50% 4320
        entries: util 33%
        bits: util 23%
        rejected: 66,72%
        LRU: 9,89
        Avg.Len. 2,78

refresh . 1440 30% 4320
        entries: util 32%
        bits: util 23%
        rejected: 67,99%
        LRU: 9,61
        Avg.Len. 2,85

refresh . 1440 80% 4320
        entries: util 33%
        bits: util 23%
        rejected: 66,47%
        LRU: 9,63
        Avg.Len. 2,78

refresh . 1440 50% 10080
        entries: util 58%
        bits: util 37%
        rejected: 41,64%
        LRU: 9,89
        Avg.Len. 2,14

refresh . 1440 50% 14400
        entries: util 83%
        bits: util 48%
        rejected: 16,47%
        LRU: 9,71
        Avg.Len. 2,00

refresh . 2880 50% 14400
        entries: util 84%
        bits: util 49%
        rejected: 15,62%
        LRU: 9,70
        Avg.Len. 2,00

 Any comments with other insights or proofs on how to get better
digests are very welcomed.

Regards,

Javier Puche.

PS: our stats and config data are at
http://www.rediris.es/si/cache/index.en.html
Received on Tue Feb 09 1999 - 11:36:19 MST