[OmniOS-discuss] Debating Swapping 2TB with 4TB drives in RaidZ2 or Create new Vol and clone

[Jim Klimov]

22 марта 2014 г. 16:15:05 CET, Bob Friesenhahn <bfriesen at simple.dallas.tx.us> пишет:
>On Fri, 21 Mar 2014, Matthew Mabis wrote:
>
>> I know the drive itself does 512b emulation but i would rather run 4K
>if theres a performance increase!
>
>Does Illumos really have a "4k" path?  It is my impression that 
>knowledge of "4k" influences offsets and allocated block sizes but 
>that otherwise things are really still done in terms of 512 byte 
>sectors.
>
>A drive which can only support I/O in 4k sectors would not be very 
>usable on most systems.


Alas (or not), that's what does happen with "honest 4k native" disks - the minimal logical io request is 4k as well as the hardware sector size, unlike the 512e drives including those which do and don't honestly report the hardware sector size which can be used i.e. to influence better alignment of system data (fs headers, etc.)

In this 4k-native case, minimal zfs block size is 4k, with some consequences in slack data overheads, fragmentation, metadata-to-data ratios, etc. There may be more visible drawbacks to such allocation on raidz than on mirrors.

In case of 512e drives, the 512b sized blocks may be used, but writes cause RMW cycles in hardware, which may reduce reliability (theoretically - just another failure mode and bug nest in logical paths; no statistics to prove practical weaknesses) and performance (once said a 30% hit for random io).

Since many OSes and FSes use 4k clusters or blocks anyway, given proper alignment to avoid RMW, they don't care or notice - they long haven't used the smaller io sizes anyway.

>
>Regardless, I can not imagine why someone would want to replace 2TB 
>drives with 4TB drives.

Limited number of disk bays? ;)

>  The resilver rate is no better with the 4TB 
>drive than with the 2TB drive so the time to resilver is doubled and 
>there are limits to what is tolerable.  I/O performance would not 
>improve and in fact it may diminish with the larger drives.  It is 
>much better to add more spindles to the pool (i.e. another raidz2 
>vdev).
>
>Bob


--
Typos courtesy of K-9 Mail on my Samsung Android