Re: *two* RAID arrays better than one?

In article <6in408$cql_at_ns.sol.net>,
Tom S. Rodman <rodmant_at_solaria.sol.net> wrote:
>We're about to install a new/better hardware RAID card in our
>quad pentium pro HP LXePro Windows NT Server (a 3 channel
>NetRAID card). Should we:
>
> - use our four 9GB drives in a single RAID 5 array
>
> - buy 2 additional 9GB disks and set up 2 separate
> RAID arrays, each with three 9GB disks
>
>The server has an informix database engine, and also
>acts as a file server to about 40-80 users. We're
>thinking of placing the NT operating system and
>the user filesystem/fileserver data on one RAID array
>and the informix database on the other RAID array.
>
>We know that there are reoccuring longish operations which
>read from the informix database and write to the filesystem.
>There are some operations that work in the other direction
>(long reads of user data on the filesystem, which then generate
>writes to the informix database).
>
>Here's an argument for for 2 separate RAID arrays presented
>by a member of our team:
>
> Here's some quick math on 1 vs 2 arrays (given a few assumptions) ...
>
> Assume an average Disk I/O rates of 6MB/sec; ex: the outside of the
> plater is typically faster than the middle and rates can vary from 7.5
> to 4 MB/sec.
>
> Assume a 4.5ms seek time; half the 9ms time listed for some disks.
>
> A 16MB file copy might be copied using 16KB blocks, thus requiring
> 1000 reads and 1000 writes.
>
> ONE ARRAY
> =========
> 16MB @ 6MB/sec => 2.67 seconds
> 2.67 second * 2 (Read & Write) => 5.33 seconds reading & writing
>
> 1000 head seeks * 4.5ms/seek => 4.5 seconds
> 4.5 seconds * 2 (Read & Write) => 9 seconds seeking for R/W
>
> TOTAL: 5.33 + 9 = 14.33 seconds for I/O
> ^^^^^
>
> TWO ARRAYS
> ==========
> 2 seeks (1 per array) => 9ms
> 2.67 seconds reading/writing => 2.67 (ie. I/O in parallel)
> 0.1 seconds (WAG) for overhead of reads & writes in parallel
>
> TOTAL: 2.779 seconds
> ^^^^^
>
> CONCLUSION
>
> ==========
> Some I/O can be drastically reduced when two arrays are used; (ex: 14
> secs down to 2.8 seconds).
>
> DISCLAIMER: The type of I/O we typically see and how RAID arrays,
> cache, etc. effect it all will vary results; we can in no way assume
> that we will these results.
>
>Can anyone poke any holes in the above argument? Are two
>RAID arrays going to give us better performance than one?

It's my understanding that you also have to account for a bottleneck in the use of one versus two controllers. Also, modern controllers order the block requests to minimize latency, smoothing out times calculated above.

>
>Please also respond by e-mail.
>
> thanks,
>
> Tom Rodman
> Johnson Controls, Inc
>
> 1-414-274-5041/fax:4400
> tom.rodman_at_jci.com
>
>
>

-- 
These opinions are my own and not necessarily those of Information Quest
jgarry@eiq.com                           http://www.informationquest.com
http://ourworld.compuserve.com/homepages/joel_garry
"See your DBA?"  I AM the @#%*& DBA!