RE: [squid-users] HELP!!!

Chuck

All that for only 1K users?

Sounds like a bit of a waste, unless of course you're seeking a very highly
available fail-over scenario; the HPs are already fairly highly available
(hot swap everything, redundant everything, depending on what you have
bought), in which case a single Alteon will also be a single
point-of-failure, as will its uplink.

Are you planning to intercept "transparently" (ie force everyone thru the
cache)? That's about the only deployment scenario that would make sense, and
even then you're highly powered.

> Please tell me what the absolute fastest model is? I
> have AceDirector layer 4 switch redirecting directly
> to squid. I have squid installed out of the box config
> standard transparent setup on a Compaq DL580 server
> Quad Xeon 700Mhz 2.5 G ram and 4 - 18 gig SCSI drives
> setup in two raid 0 for speed with RedHat OS.
>
> I Have 3 of these quad servers and am thinking of
> parent-child but need advise as to over all plan. This
> is being setup with 1000 web users in mind.
>

I'd use that kind of power for a small country, but if you have it, it
should absolutely smoke. I question your choice of Linux, but it may be OK,
just steer away from the ext3 FS, and definitely no journalling (you
wouldn't, right??). I've had good success with FreeBSD v4.x with Squid on
HP, goes like stink and few noticable FS peformance problems, but them the
right choice of FS under RH may work OK.

As for peering the cache system, I have mixed feelings on this one. I have
rarely seen deployment scenarios whereby the cacheable content mix present
on a system of caches performed better through ICP than through refetching
from the source (and hence redundant objects present across the caches)
except for the longest lived objects, which are usually small-to-average in
size anyway, unless you are at the bottom of a really slow uplink.

I have, at times, choked up transit links with ICP overhead, back in the
days where these links were small, but then I ran my system in distributed
farms across a geographically dispersed topology ... it sounds to me like
you're clustering these boxes at one point?

Need more data ...

J.
Received on Mon Nov 21 2005 - 22:04:36 MST