Re: [squid-users] Squid Performance

From: Brian <[email protected]>
Date: Thu, 21 Nov 2002 12:21:31 -0500

On Thursday 21 November 2002 11:19 am, Simon Mullis wrote:
> Hello All,
>
> I would be very interested to understand the key performance bottlenecks
> that one would seek to avoid when commissioning a Squid service. I am
> primarily concerned with sustained HTTP throughput.
>
> What elements of a chosen hardware platform would have the most effect
> on performance for Squid running on Intel based Linux system?
>

> Number of CPUs?

The squid core runs as a single process, so even a second CPU will have
minimal benefits. Any more is just burning money.

> Memory I/O?
> Front side bus speed?
> P4 v. P3 v.

These are not really a factor. Squid is almost purely IO bound.

> CPU 2nd level Cache?
> CPU clock-speed?

These will make a difference, but no need to go crazy. A large-cache
(512k, I think) P3 should be fine.

> Mounting the disks Asynchronously?

This should be the default but yes, write-behind cache is critical.

> Disk subsystem throughput?

Critical. If you read the archives, you will see a lot of setups with 5-6
partially used SCSI drives. Seek time will really hurt your hit
performance. No RAID, just many cache_dir lines.

If you're on a budget, consider a 3ware IDE RAID card with each disk
independent.

> Memory capacity

I added this one. The best way to minimize seek is to have your working
set entirely in RAM. Go for a board that supports 4gb, though you may
only need 2gb or so for now.

> I am aiming for a maximum throughput of greater than 20Mbit/sec per
> cache.

The hardware you will need for this depends a lot on the usage pattern you
receive.

        -- Brian
Received on Thu Nov 21 2002 - 10:21:03 MST

This archive was generated by hypermail pre-2.1.9 : Tue Dec 09 2003 - 17:11:29 MST