Re: [squid-users] Performance tuning Squid box for ISP traffic

From: Joel Jaeggli <joelja@dont-contact.us>
Date: Thu, 9 Dec 2004 11:09:39 -0800 (PST)

On Thu, 9 Dec 2004, Henrik Krohns wrote:

> On Thu, Dec 09, 2004 at 03:58:07PM +0800, Ow Mun Heng wrote:
>> On Thu, 2004-12-09 at 13:52, Lucia Di Occhi wrote:
>>>> Do not use RAID.
>>>
>>> Do you mean SOFTWARE RAID? Hardware striping will definately improve
>>> performance.

To address this assertion, You will frequently find linux software raid to
have superior performance to hardware raid subsystems especially if you
have lots of bandwidth available to your controllers. that something to
need to benchmark for your self but when you compare that rate at which
you processor (a 3GHZ xeon for example) can xor data which is order of
4GB/s vs the risc core on your raid controller which is probably order of
200-300MB/s, that probably not an issue on a 32 bit 33mhz pci bus but it
can be with 2 x 64bit 66mhz buses or more.

>> Frankly I'm not sure. Most ppl on the list does not encourage using
>> RAID. I only run the cache on 1 SATA disk.

The performance isssue with raid is when you have cache objects larger
than your stripes chuck size you are serializing requests across multiple
disks. That means you're moving two or more spindles to read or write the
same object when you could move only one. The fastest hard-drive you can
buy (maxtor atlas 15kII) has a read-service time (access plus rotational
latency) of around 5.5ms which means that even under ideal conditions it
can only access the media around 181 times a second. Add in the additional
overhead of additional writes on raid 1 or 1+0 (twice as many) or the
xored parity on raid 4 or 5 (about 1/8 more in most cases) a you have more
overhead...

The question becomes is it faster to stripe say 3 x 15k rpm disks or have
three seperate cache dirs one on each disk. squid performance is really
measured on a basis of number of objects per second (or transactions if
you're a database person) not throughput bytes in-out. This issomething
you may need to benchmark on your hardware in order to be shure of what
you're seeing.

The issue of squid disk io performance and tuning gets a lot of treatment
in Duane's excellent squid book in chapter 8 advanced disk cache topics,
and appendix D filesystem performance beenchmarks. ISBN 0-596-00162-2

> People should be more specific than just "dont use raid!". Modern hardware
> raid-boxes with GB's of cache are blazingly fast even on RAID-5. Ofcourse
> you shouldn't spend such money for cache which you can lose without problems,
> but if you have some free space on some raid-box why not use it..
>
> Cheers,
> Henrik
>

-- 
-------------------------------------------------------------------------- 
Joel Jaeggli  	       Unix Consulting 	       joelja@darkwing.uoregon.edu 
GPG Key Fingerprint:     5C6E 0104 BAF0 40B0 5BD3 C38B F000 35AB B67F 56B2
Received on Thu Dec 09 2004 - 12:09:55 MST

This archive was generated by hypermail pre-2.1.9 : Sat Jan 01 2005 - 12:00:02 MST