ZFS performance bottlenecks: CPU or RAM or anything else?

[Brandon J. Wandersee]

Alex Tutubalin writes:

> On 5/17/2016 3:29 PM, Daniel Kalchev wrote:
>
>> Not true. You can have N-way mirror and it will survive N-1 drive failures.
> I agree, but 3-way mirror does not looks economical compared to raidz2.

If you're already planning for multiple simultaneous drive failures,
"economical" isn't really a factor, is it? Those disks have to get
replaced regardless of the redundancy scheme you assign to them. ;)

Whether the concern is performance or capacity, mirrors will offer the
most flexibility. Increasing either the performance or capacity of a
RAIDZ pool necessitates either replacing every disk in the pool or
doubling the number of disks in the pool, all at once. Mirrors allow you
to grow a pool and increase/decrease redundancy asymmetrically. True,
four disks in a two-mirror stripe will see you restoring a backup if one
disk from each mirror dies, but (arguably) six disks in a two-mirror
stripe offer both better redundancy and better performance.

Speaking strictly about performance, RAIDZ performance is pretty much
fixed, while mirrored performance will (I believe) increase slightly as
you add disks and increase greatly as you add vdevs.

-- 

::  Brandon J. Wandersee
::  brandon.wandersee at gmail.com
::  --------------------------------------------------
::  'The best design is as little design as possible.'
::  --- Dieter Rams ----------------------------------