Optimizing performance with SLOG/L2ARC

[Bob Friesenhahn]

On Wed, 19 Aug 2015, kpneal at pobox.com wrote:
>> Use mirrors if you can afford it.
>
> Mirrors are less safe than raidz* unless you have enough drives in the
> mirror. Look up the failure prediction calculations that were done on one
> of the zfs lists last year. (The relevant keyword might be "MTTDL".)
>
> Mirrors also have about the same write performance as raidz*. But reads
> from mirrors scale excellently with the number of drives in the mirror.

With traditional hard drives, mirrors offer more write performance 
than raidz because they consume fewer precious drive IOPS (e.g. 5X, 8X 
less) and because the pool will have more vdevs (supporting more 
simultaneous activities).  With single large sequential writes, raidz* 
likely matches performance with mirrors.  Performance with large 
sequential writes is not usually the problem when someone says that a 
server feels "sluggish".

With today's large disk sector sizes (4k, 8k), the amount of space 
lost to mirroring (compared with raidz*) is not as much as it used to 
be.

It is necessary to start at raidz2 in order to exceed the MTTDL with 
simple mirrors.

Bob
-- 
Bob Friesenhahn
bfriesen at simple.dallas.tx.us, http://www.simplesystems.org/users/bfriesen/
GraphicsMagick Maintainer,    http://www.GraphicsMagick.org/