Re: [SQU] recommended hardware

From: Joe Cooper <joe@dont-contact.us>
Date: Wed, 01 Nov 2000 15:18:19 -0600

Christian Schmit wrote:
>
> Hi,
>
> We plan to setup a transparent caching service
> using Squid 2.3 with WCCP. Testing has been
> successful so far.
>
> Now we need to buy the production servers for this.
>
> I plan to start with 2 servers. They should be able to handle
> together around 10 Mbit/s of www traffic. They should scale by
> adding more servers to 45Mbit/s or even more.
>
> I got recommended that it would be better for scalability/performance
> to install more smaller servers than some big servers.

Yes, quite true.

> What is a recommended PC for this type of application?
>
> CPU ?

Depends on the compiler options you use in Squid. Because you're using
2.3, you can't use async i/o because of instability and low hit ratio
when under load. So CPU isn't really a big factor. Even with three or
four disks (more is mostly wasted on Squid currently), you won't be
loading your processor down terribly much.

If you plan to use async i/o in the future (Henrik pretty much has it
fixed in 2.4 as far as I can tell...I'm starting heavier stress testing
today), then give yourself about 200-300MHz per disk that you plan to
use.

> RAM ?

A bunch. Per ~30GB of cache I recommend 512MB of RAM, if not
more...otherwise performance and stability suffers.

> IDE/SCSI?

7200 RPM IDE provides better price/performance currently. SCSI is
faster. Depends on your needs. Since you've already decided that
multiple balanced caches is OK, then several small boxes with dual 7200
RPM IDE drives is your cheapest option. Maintenence of fewer larger
caches is probably easier, however, so maybe SCSI is worth the extra
money. (A single big SCSI box can probably handle your current load,
depending on OS and tuning.)

> HD Size ?

Doesn't matter. Most modern hard disks are too big for a Squid based
cache. Since the affordable RAM limit is usually 768MB, that means you
shouldn't operate with more than ~38GB of cache_dir space per box. Even
that is probably pushing it.

> Cache size per HD ?

Less than 85% of the partition filled (and keeping it under the above
mentioned limitations based on RAM capacity).

> How many HD per PC?

2 for non-threaded and non-diskd squid. 3-4 for async i/o or diskd.
Beyond that the law of diminishing returns rears it's ugly head.

> HD should be formatted with what "block size" 1K, 2K, 4K etc...

ReiserFS mounted 'noatime,notail' if you're running Linux. Otherwise,
2K seems a good number in benchmarks, but 4K isn't much different. I
wouldn't worry over block size too much. Mount noatime, regardless of
FS, of course.

> Any HD models that can be recommended ?

IBM's 36LZX model SCSI 10k drives are fast and cheap, with good seek
times (4.9ms), their IDE drives are not so good however. Seagates SCSI
drives are slightly faster than IBM, but cost quite a bit more. Maxtor
and Seagate seem to make the fastest seeking IDE drives for Squid load.
(Quantum's are faster than Maxtor, but I've had several reliability
problems with them recently so I avoid them.)

If you're running Linux (even if not, it may be helpful to some degree),
take a look at a tuning article I've written that discusses most of
these topics:

http://www.swelltech.com/pengies/joe/squidtuneup/t1.html

Hope this helps.
                                  --
                     Joe Cooper <joe@swelltech.com>
                 Affordable Web Caching Proxy Appliances
                        http://www.swelltech.com

--
To unsubscribe, see http://www.squid-cache.org/mailing-lists.html
Received on Wed Nov 01 2000 - 14:13:37 MST

This archive was generated by hypermail pre-2.1.9 : Tue Dec 09 2003 - 16:56:13 MST