Re: [SQU] Disk access as Squid bottleneck?

At 07:46 PM 11/19/00 -0500, Steve Snyder wrote:
>My Squid (v2.3S4) cache has plenty of CPU and RAM available to it. Also,
>the hard disk on which the cache is stored has twice transfer rate of the
>100Mbps network that the data crosses. Given these facts, am I correct in
>thinking that disk access time is the bottleneck for my cache's throughput?

You mentioned the access time (or seek rate) of your disk. Yes you will
probably get 60MB/s out of your disk but you can only hope to approach that
when you access a file that is sequentially stored on your disk (which is
never the case on a server) and when the file is long enough to ignore the
8.5ms access time.

The problem is that you will get a chunk of data. Then the head has to seek
for another track. You'll get another chuck etc etc. The average seek time
is 8.5ms. Now calculate again and assume the data is not sequentially
accessed/stored. That means that the IDE data transfer is randomly
interrupted and no data is transferred for an average time of 8.5ms.

The solution: Use more disks to balance the load. SCSI is generally better
when the load is balanced over more disks. (I run a 12G cache divided over
4 disks and that works fine on a Sun Ultra1-140, ATM OC-3 NIC, 300 users
and 1Mbps WAN)

>My hard disk is a very good IDE drive, but it has an IDE drive's access
>time: 8.5ms. With sub-5.0ms SCSI drives on the market, I wonder if it
>makes sense to get one as a replacement for my current drive. I'm willing
>to spend the money, but I want to feel confident that the reduced access
>time will really be noticeable as improved cache performance.
>
>Any thoughts on this topic?
>
>*** Steve Snyder ***
>
>--
>To unsubscribe, see http://www.squid-cache.org/mailing-lists.html

--------------------------------------------------------------------
Marc van Selm
NATO C3 Agency
Communication Systems Division, A-Branch
Tel: +31 70 3142454
E-mail: marc.van.selm@nc3a.nato.int (PGP capable)
Tactical ATM: +31 70 3148183 / 71183
--------------------------------------------------------------------
Private: selm@cistron.nl, selm@het.net, http://www.cistron.nl/~selm

--
To unsubscribe, see http://www.squid-cache.org/mailing-lists.html