Re: comments?

From: Richard Archer <rha@dont-contact.us>
Date: Thu, 16 Dec 1999 10:36:17 +1100

At 7:01 +1100 16/12/99, Jon Mansey wrote:

>Below is my cachemgr info page at our peak time today.
>
>Im a little disappointed by the byte ratio, any comments?
>
> Request Hit Ratios: 5min: 36.7%, 60min: 36.5%
> Byte Hit Ratios: 5min: 15.9%, 60min: 18.4%

These figures are only samples at a specific time. I have found them to
be quite volatile. Try running calamaris over your access.log, and see if
it gives you different results over a longer timeframe.
http://Calamaris.Cord.de/

You can increase the hits-by-byte at the expense of hits-by-request by
using 'replacement_policy LFUDA' and increasing the maximum_object_size.

You can also increase hits-by-bytes somewhat by judicious use of
'reload-into-ims' directives in the refresh_patterns:

refresh_pattern -i \.gif$ 10280 90% 43200 reload-into-ims
refresh_pattern -i \.jpg$ 10280 90% 43200 reload-into-ims
refresh_pattern -i \.exe$ 10280 90% 43200 reload-into-ims
refresh_pattern -i \.mov$ 10280 90% 43200 reload-into-ims
refresh_pattern -i \.mpg$ 10280 90% 43200 reload-into-ims
etcetera

You can generate from your access log a report of bytes per file extension
with: calamaris -n -t -1 -O

This is especially effective when used in combination with the LFUDA
replacement policy (i.e. you have to have the object cached before there
is any point ims'ing it, and LFUDA is best at keeping large static objects
in-cache). Caveat: this violates the HTTP protocol.

And of course the best way to increase the hit rate is to hold more data
in the cache... you are only holding 4 days' worth, which seems a little
low to me. There is certainly a point of diminishing returns though, and
you'll have to determine that point for your particular environment.
I imagine your diverse customer base will create usage patterns that would
benefit by having more storage on-line.

I certainly think holding 8 days' worth of data is worth while, as there
is likely to be sites that get hit in weekly bursts (like sports sites
that get accessed mainly on game days).

Does anyone know of a log analysis tool that can help determine the optimal
cache size? I imagine this could be done by processing the access.logs, and
working out how frequently recently-expired objects are re-requested.

...Richard.
Received on Wed Dec 15 1999 - 16:44:21 MST

This archive was generated by hypermail pre-2.1.9 : Tue Dec 09 2003 - 16:49:54 MST