RAID 10 Sanity Check

[sechs]

For a four-disk RAID 10 array on two channels I want to:

a. Stripe down channels and mirror across

-or-

b. Mirror down channels and stripe across


*Extra credit points will be awarded for best reason why a or b is correct.

 

[Mercutio]

You're mirroring first, then striping. As the name implies: 1+0.
If you were striping and then mirroring, failing even one drive would fundamentally damage the array - the stripe would be gone, at least, and you wouldn't have a functional "01" array any more.
In a mirror + stripe, there's at least a possibility that the array could surivive multiple drive failures. That's why we like it better.

 

[Howell]

And to answer the question I thought you were asking, you want the striped drives on different channels because you will be reading/writing to them at the same time.

The OS should be able to queue up reads and writes to the mirror more easily.

 

[sechs]

Since Merc read more into the possibilities than was there, I'll go with Howell's answer.

My theory on this (and I may be wrong) is to go with (a) because the same data will always be going (being written) to the mirror pair, I'd want them on different channels; otherwise the channel will get clogged-up with sending write data to one drive and the other of the mirror pair. I do, however, see the point of stiping across the channels for throughput.

Does the case need to be more specific to delinate whether a or b would be (more) adventageous?

 

[ddrueding]

Since Merc read more into the possibilities than was there, I'll go with Howell's answer.

My theory on this (and I may be wrong) is to go with (a) because the same data will always be going (being written) to the mirror pair, I'd want them on different channels; otherwise the channel will get clogged-up with sending write data to one drive and the other of the mirror pair. I do, however, see the point of stiping across the channels for throughput.

Does the case need to be more specific to delinate whether a or b would be (more) adventageous?

I think you should give yourself extra points, but your theory applies more if you feel the majority of your writes will fit in a single stripe, otherwise you're screwed either way.

 

[sechs]

Okkay, I see. If you have to write across channels (through multiple stripes), it doesn't matter anyway.

 

[Howell]

Okkay, I see. If you have to write across channels (through multiple stripes), it doesn't matter anyway.

Yeah, That's why I was saying it was easier for the OS to delay writing to the mirror than delay writing the next stripe. Although a relative measure, the stripes are not really that big and it is likely to need to write multiple stripes.

You do realize only one drive on an IDE channel can be active at a time, right?

 

[sechs]

Yes. That's why I'm talking about SCSI.

Does a hardware controller make a difference? Since the data would be sent once to the controller, it would duplicate it to the mirrors.

 

[ddrueding]

Before I wasn't sure you were talking about SCSI, so I was hedging my answers (you weren't really specific). How many drives are you planning on using? If the # of drives per channel * the STR of the drive =! the speed of the channel (160 or 320) than it matters not. If it is an issue, the above arguments stand.

 

[ddrueding]

=!

should be <

 

[LunarMist]

Why not simply try it both ways?

 

[sechs]

That would involve having the appropriate hardware....

If the # of drives per channel * the STR of the drive =! the speed of the channel (160 or 320) than it matters not. If it is an issue, the above arguments stand.

That makes sense. Presuming, for a moment, that the maximum amount of data that the drives on a channel could put out was greater than the speed of the channel, then striping across would be the choice?