[squid-users] suggestions for high-end dual-CPU config on Linux? (+tuning of current conf)

Hello,

I have read the list archives but although this is a common question,
I have not found any information precise and up-to-date enough for my
taste :-)

We do transparent caching for a one-way satellite Internet access
service. This means a *lot* of traffic! Currently the caching is
load-balanced over two servers:

- one is P4 @ 1.8Ghz
- one is P3 @ 1.3Ghz

They run the Linux 2.4.19 kernel, Squid 2.4-STABLE7 over a ReiserFS
filesystem with noatime,notail.

Both servers have 1Gb of RAM, two or three Ultra-160 SCSI disks
hooked to an Adaptec controller. The cache is split over these 2 or 3
disks using software RAID0. It is about 25Gb big.

Each server has 2 LAN interfaces:

- one supporting the traffic to/from customers
- one supporting the traffic to/from the Internet

These are Intel Etherexpress PRO/100 and 3COM 3C905 interfaces.

These two boxes are quite overloaded now. Load averages over 10 are
not uncommon and response time gets quite slow at peak hours. It looks
to me that the interrupt rates from the LAN interfaces and disk I/O
are the main bottlenecks here, but I'm not that sure.
Logs have been completely disabled for now to alleviate the problem.
It has helped somewhat.

I understand that some tuning could be needed there (advice welcome).
Especially the RAID0 might not be that a good idea. Having separate
cache dirs might be more efficient.

Would upgrading the LAN interfaces to something better (what?) help?

If anyone wants to help, I can provide detailed stats of these servers
upon request.

Anyway, it looks obvious that these boxes need a hardware upgrade soon
because the customer base is growing fast.

Can anyone provide good experience with a Linux-based dual-P4
configuration and hardware RAID controllers, plus "smart" LAN
interfaces? Details of configuration are quite welcome.

Thanks in advance,
Greets,
_Alain_

--
"I've RTFM. It says: `see your system administrator'. But... *I* am  
the system administrator"
(DECUS US symposium session title, author unknown, ca. 1990)