Raid Group size 8, 16, 28?

[jasonczerak]

So, I've search and never ran into why a RG size of 15 or 16  vs for example 28.     I realize that a RG size of 28 the rebuild time is longer, but the space is more efficiently used. however a RG size of 8, rebuild time is quick, over all the data is better protected, however your wasting disks.

Can it be true the number 16 is purely just a happy medium size to use for RGs?   Maybe dropping down to 15 with 300G disks to get a little closer to the 16TB limit?

Or is there some sort of performance penalty going up to 28 disks because of how the algorithms work?

[__frostbyte_9045]

I second the fact that it's just a target best practice.  I deviate some in order to match fully populated disk shelves to raid group size.  For example, when we started, we added one DS14 disk shelf at a time you may want to use a RG of 13 to facilitate a spare per shelf to prevent RG's of 3.  On the other hand with a 24 disk shelf, I'd have to change my strategy to maximize useable space with out going to some crazey number (say 23 disk RG) based on the expected shelf count.

[druddle00]

For SATA 1Tb disks, I generally use a RG size of 12 if I am configuring a 23 disk aggregate (i.e. max space).  Give this thread a read - http://communities.netapp.com/thread/1587

Dave

[evilensky]

JasonCzerak wrote:
Or is there some sort of performance penalty going up to 28 disks because of how the algorithms work?

The resiliency guide is silent on issues of performance.  Pg. 22 says

Smaller RAID groups

Use smaller RAID groups for faster reconstructions and reduced risks during reconstruction.

Page 11:

NetApp recommends using the default RAID group sizes when using RAID-DP.

Page 4:

Use the default RAID group size when creating aggregates or traditional volumes.

All disks in an aggregate are supposed to participate in IO operations.  There is a performance penalty during reconstruction as well as risks; "smaller" RG sizes are meant to minimize both.

There is a maximum number of data disks that can contribute space to an aggregate for a 16TB aggregate composed entirely of a give disk size, so I've seen RG sizes deviate from the recommended based on that factor (You don't want/need a RG of 2 data+2parity just to add 2 more data disks to an aggr....). Minimizing losses to parity is not a great solution to any capacity issue.

my $0.02.

Message was edited by: evilensky

[jasonczerak]

Thank you, I think that for the most part clears things up. We'll just stick with 16disk RG's until we find a funky need to add one or two disks.

The interesting part tho that I tend to over look at times is the basic capability of the individual controlers. So looking at this from an IO stand point is pointless with a 3160. If you have a 4 disk RG mixed in with a 16disk, that should pose a problem in the stripe of data across the RG's, but not a 15 disk in a 16.The IOs the disks are capabile far outweight what the contorler can dream of handling. once you get 8 or so shelves in play.

[mcope]

An enterprise account I'm familiar has been using NetApp storage since F300 days and they have tested all types of configurations and have found performance starts to flatline after 16 disks.  I think the most convincing proof that 16 is the sweet spot is the results on spec.org.  NetApp tests using 16 disk RAID groups.