Re: Creating a Database on RAID 5

Ed Stevens wrote:

> My adv. frankly donot use anything _BUT_ RAID 5, all others have a cost
> penalty . . .
>
> Of course, my _real_ point is that the choice of RAID configuration --
> or lack of -- depends on the situation. Yes, all else being equal
> RAID-5 has a performance hit on write operations. BUT . . . "all else"
> is never equal. RAID implemented in hardware with a good caching
> controller will improve performance over an OS implementation. More
> spindles improve the performance through increased parallelism,
> especially on read operations -- and given the application mix, this
> might acutally IMPROVE overall performance. Perhaps the application is
> simply not very write intensive when compared to the I/O capacity of the
> disk sub-system. In this case a slower write operation may very well
> not even be detectable, especially to the end user.
>
> While I don't have the benchmarks, I'd be willing to bet that the number
> of physical channels you can spread your files (and thus, your I/O)
> across has more of an impact on real throughput/response time than the
> RAID level.

I mostly agree with this. The advantages of having RAID5 for mission critical data cannot be underestimated. I had an experience about a year ago when one of the disks in the array had a gross media failure and was delighted when I found out that even without taking the server down we could just pull out the failed driver, insert a new one and all without the users even knowing that anything had happened.

I do recommend that the online redo logs be placed on RAID1 and there is a white paper on the Oracle web site which explains the reasons for this, but I have also been in situations where I did not have this choice and had to put the online redo longs onto the RAID5 array but the server and disk array were so high-end that I couldn't point to a single statistic which I could use to support the contention that performance was hurt by this - in fact, the system was so overpowered that I had to work really hard just to get cpu availability to drop below 90%. Received on Wed Oct 25 2000 - 14:07:55 CDT