Re: [SQU] Hardware RAID0 or 2 independ disks

Fabien Salvi wrote:

> Hello.
>
> I've posted a mail 2 weeks ago about problems with squid.
> They are now resolved, thanks to the help of Joe Cooper.
> So, I repeat my material :
> PII-450 Mhz Bi-pro
> 512 MB RAM
> 9GB HD SCSI (RAID1 hardware with external controller)
> NIC 3com 905b
> Debian GNU/Linux 2.2 (with glibc 2.1.3)
>
> squid :
> Squid :
> squid 2.3.4 with squidGuard redirector (15 redirectors started)
> Mem config in squid to 128
> Cache_dir to 3200 MB
> Reiserfs filesystem (r5 hash)
>
> In fact, I have 25 to 35 requests/s
> (What is the difference from the number returned by squid manager client
> on 5min ==> squidclient mgr:5min server.http.requests and what I obtain
> from mrtg in cacheServerRequests that is the double ?)

>
> I am going to add 2 disks.
> I have 2 solutions :
> 1) Using my RAID external controller (Mylex DAC960SX), I put the disks
> in raid0, so I shoud have high performance
>
> 2) Using my RAID controller, I can put the 2 disks in independant, so I
> configure 2 different cache_dir
>
> You must know that I can't use disks outside my raid controller...
>
> What is the best for high performance and increased load (should be more
> than 50 requests/s in a few months) ?
>
> Thanks in advance for your help !

50 requests/s is easy on even a non-async compile of Squid (we get more
than that from , and adding a couple more SCSI disks will only increase
your performance (assuming you don't go into swap...so lower your
cache_mem to 64MB, probably).

Either RAID 0 or separate cache_dirs will be fine for your workloads,
although I usually go with more traditional controllers and separate
cache_dirs. The performance difference would be felt if you were
hitting your cache harder than 50/s, because with striping you have
parity time and you're chopping up very small files (average 11k
probably) into potentially smaller chunks (8k, probably)...you end up
having the latency of the slowest disk that has a piece of the file on
every request rather than the latency of a single disk. Because of
this, since you have a choice, go with separate cache_dirs. You'll
still get the benefit of the big RAM cache on your Mylex, and none of
the negatives of striped disks.

There will probably come a time when Squid uses larges block writes and
reads, and striped disks will probably work great at that point, but
until then it's probably most effective to avoid them.

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