wojtek@3miasto.net wrote:
>>I am thinking of setting up squid for caching having worked with it in the
>>past, but Im unclear about some sizing issue for a large installation.
>>We are a broadband (cable modem) based ISP with a some dialups too. My
>>immediate need is to service around 50 - 60 http requests per second.
>>
>
> use one separate drive for this (count about 100 req/drive)
>
>
>>This got me to the big figure of 4,320,000 per day, which I didnt not find
>>in the JANET artice by Martin Hamilton on the FAQ. Also with an average 8k
>>object size, I gather id need 32GB hard disk and 512MB ram minimum.
>>
>
> why 512MB RAM?
>
> 256MB is enough for 32GB cache.
No, it's not. Nowhere near it, in fact. Especially not at the loads he
is seeing. In fact, I would be uncomfortable rolling it out with less
than 768MB for 32GB of cache. (The old rule of thumb of "10 MB for
every GB of cache_dir plus cache_mem plus whatever the rest of your
system needs" is a very good rule to follow when configuring Squid. I
like to add a little RAM on top of that, just because it makes
performance so much snappier for real world conditions.)
>>My questions:
>>
>>At what point do I need to consider clustering ? (Note the above is at one
>>single physical location.)
>>
>
>
> but 2 fast 7200 IBM IDE drives (30GB), 512 MB RAM, make mirrored 2GB
> partition for OS, logs, small swap etc and 28GB for squid.
This will handle about 110 reqs/sec, and about 10Mbps total throughput
(or about 7Mbps actual uplink bandwidth) in a Linux and Squid+async i/o
configuration. On FreeBSD with DiskD this will yield about 80-90 reqs/sec.
> don't forget to mount -o noatime , use 1 inode/8kB and 2kB blocks in linux
> and 8/1 blocks in *BSD.
>
> is you can choose system, choose *BSD (i use netbsd). linux isn't usable
> on high load
Silly anti-Linux FUD. Linux and Squid+async i/o on ReiserFS is the
fastest Squid configuration you'll find, and our boxes configured this
way are installed in a number of very high load environments with /no/
stability problems. We measure our client system uptimes in months, not
weeks or days. (Also, if more evidence is needed, we nearly matched
Duane's FreeBSD box at the last cacheoff with approximately half the
hardware--the results are public, along with system specs.
http://www.measurement-factory.com )
I won't tell folks to use Linux over FreeBSD...you should use what
you're comfortable with. But don't spread misinformation, just because
you prefer one OS over another.
--
Joe Cooper <joe@swelltech.com>
Affordable Web Caching Proxy Appliances
http://www.swelltech.com
Received on Wed May 09 2001 - 15:22:53 MDT
This archive was generated by hypermail pre-2.1.9 : Tue Dec 09 2003 - 16:59:54 MST