2009/8/4 Hery Setiawan <yellowhat89@xxxxxxxxx>:
> maybe in his mind (and my mind too actually), with big mem_cache the
> file transfer will be transferred faster. But using that big for me
> it's too much, since I only have 4GB of RAM and thousand workstation
> connect with my squid.
The squid memory cache doesn't work the way people seem to think it
does. Once objects leave the memory cache pool they're out for good.
The rule of thumb is quite simple - keep cache_mem large enough to
handle in-transit objects and a few hot objects; leave the rest to be
available for general operating system disk buffer caching. Only
deviate from this if you absolutely, positively require it.
This will occur when your workload has a lot of small objects which
you frequently hit. Hack up or download something to generate a
request size vs {hit rate, byte hit rate, service time, cumulative
traffic} to see exactly how many tiny/small objects you're getting a
hit off of.
If you have a very small set of constantly "hot" traffic that will fit
in memory, up cache_mem. But be aware of the performance repercussions
if the "hot" traffic leaves cache_mem and stays on disk.. :)
If you have a set of "hot" traffic that moves over time, upping
cache_mem may not help.
2c,
Adrian
