dinsdag, maart 22, 2005

The best Raidlevel (Part 2)

So as it turns out, Raid 5 is a very good Raidlevel, but it is not very good at writing. Raid 1+0 is very good, but it is rather expensive. Is that all there is to it ? No. If you have read one of my previous articles, you might have read that a harddrive’s speed is influenced by Latency and Seek. If you combine that story with this particular story, then you might find out there is more to it.

The two main ways of communicating with a disk are Sequential and Random. In a sequential environment, large chunks of data are being transferred (e.g. : 64K or 128K or bigger) This occurs most on file and print servers. In a random environment, larege amounts of small chunks of data are being transferred to the disk (e.g. 4k) This occurs most on a database server.  If you think a little further on this, you might consider that Raid 5 has to do about twice the work for writing a chunk of data compared to Raid 1+0 (Raid 5 has to read existing data, read existing parity, calculate the parity, write new data, write new parity).

In that case you might come to the same conclusion as Database manufacturers and database gurus. Raid 5 is not a good thing to combine with a database. It will flood the Raid 5 with small chunks of data and that poor Raid 5 has to do twice the work a Raid 1+0 has to do at writing.

A clear conclusion is : Avoid Raid 5 in a database environment. But most people forget that Exchange is also a database.

My recommendation : If you are going to install a server, and you are considering to install 3 disks in Raid 5 and put your exchange server on it — why not consider a little further, add 1 extra disk, use Raid 1+0 in stead. The result is remarkable : Twice the speed, Twice the fault tollerance, twice the speed at reconstruction, more than twice the speed during a disk failure.

Geen opmerkingen: