Need ICH9R Raid 5 setup advice

[Charles W. Stumpfel]

I'm trying to get my new raid 5 array happy before i start dumping gig after gig onto it.

Hardware:
Gigabyte ga-p35-ds3r
q6600
3 WD 500gig AAKS Drives - miserable in RAID 5
2 WD 74gig Raptors - Happy in raid 0 (128k stripe)

I want both arrays on the ICH9R. The problem is i can't get the RAID 5 to perform acceptably. The first go around i tried a 64k stripe size and 4k cluster size. Ran hdtach(i know its a poor benchmark for raid 5...) and got poor seq read ,67MB/s. So i figured i'd re-do the whole thing after doing some research and tried 32k stripe and 64k cluster. Well reran hdtach and got 55MB/s read. Said to myself, screw it, im sticking with this set-up. Starting copying roughly 100GB of data(100k-1gb media) onto the array. 2 hours later it was only 1/3 copying and i canceled it. Was getting roughly......10-15mb/s write speed from what i can tell. And from what i can tell WB cache was enabled.

I'd like this to be my final re-do. Its going to be a media only drive with files ranging from 100k-5gb. I've found too many conflicting theories on the net about cluster/stripe ratios and whether large sequential files need big/small stripe sizes. My next try was going to be 128k stripe with either 4k cluster or 64k cluster.

Any input would be hugely appreciated considering it takes several hours to re-format/test each of these set-ups.

Thanks in advance,
chuck

 

[shade_star]

Im not familiar with these cards but if its Intel based, you have to enable write back cache.

 

[stueyG]

my understanding is that onboard raid 5 can be pretty slow because the CPU is being used too generate the needed parity information - so if your doing something processor intensive this can bottleneck the system.

 

[Madwand]

128k is a bad choice for Intel RAID 5 from what I've seen. It's just an implementation detail -- there's no reason for this that I understand. I suggest 32 or 64k for a 3-drive RAID 5 array.

32k comes from dividing a common 64k access among 2 data drives. 64k is that same magic number again, used as a default because lots of people use it as a default, so there could be some tuning for it.

You may need to enable write back caching in order to improve write performance.

With decent caching, the size shouldn't make a big difference (except for the 128k anomaly).

You could be seeing two different effects:

1. HDTach is not the greatest tool for benching RAID arrays.
2. Write performance will be poor unless write back caching is enabled.

ATTO is a quick & dirty tool which will do read + write testing with varying access size. Here's some sample performance & setup information, using ICH8DO for 4-drive RAID 5 using 64k stripes.