Re: [squid-users] benchmark test

From: Joe Cooper <>
Date: Tue, 19 Feb 2002 02:46:41 -0600

Comments below:

Bgs himself wrote:

> On Mon, 18 Feb 2002, Marc Elsen wrote:
>> You may get some idea's at :
> That's a nice place if I'm starting from zero, but I need to make some
> actual calculations like: How will TPS increase with
> memory/CPU/cachesize/disknumber/CPU? Disk types compared to each other
> (IDE/various SCSI falvours/IDE RAID0/SCSI RAID0)?
> These test are mainly helping someone to find the optimal performance from
> a certain system. I need at least rough estimates about performance
> depending on hardware upgrades...

Why not perform your own tests? Polygraph is free for the downloading.

> Something like:
> PI-133MHz/64MB/512MB-IDE -> 30TPS

Probably. If using aufs.

> PII-333/128/1GB-IDE -> 60TPS

Probably. If using aufs.

> Dual PIII-800MHz/1GB/4GB-SCSI -> 200TPS

Probably not. Even if using aufs. You'll need 4 10k or 3 15k disks to
get up to 200TPS, and faster processors than that. Oh, yeah, you'll
have to run two Squids with different configuration files to take
advantage of the dual CPU. This system, with only a single SCSI disk,
would do 100 on a good day. Give it two to four disks and you'll scale
up to 160 or 180.

> Athlon-XP 1.5GHz/2GB-DDR/8GB-SCSI RAID0 -> 500TPS
> for the same software setup.

Fantasy. This would take, at least, dual Athlons (as fast as money can
buy), 4GB of DDR, and 8 15k disks...and even then, it's probably
fantasy. The CPU would be the bottleneck in this scenario--when running
aufs CPU usage is exceedingly high, but you can't scale at all without aufs.

This system, as you've defined it with a single disk, will also max at
100 on a good day--with 2-3 disks will push nearly 160-180, as above,
most likely. Again, the bottleneck is CPU and you aren't giving Squid
any more CPU than the previous dual 800 version. Forget the RAID, even
RAID0. A hardware RAID controller does have extra buffer RAM that may
be of value--but actually using RAID of any sort is generally
counter-productive. Squid will disperse requests roughly evenly across
all disks, without the overhead and object breakage of most types of RAID.

Of course...maybe you're talking about website accelerator performance,
in which case all these numbers go out the window. Squid is much faster
from a RAM only workload, so your high numbers for those last two boxes
get to be more realistic.

Joe Cooper <>
Web Caching Appliances and Support
Received on Tue Feb 19 2002 - 01:47:13 MST

This archive was generated by hypermail pre-2.1.9 : Tue Dec 09 2003 - 17:06:26 MST