[zfs-discuss] recommendations for a 50 TB zpool?

Richard Elling richard.elling at richardelling.com
Sun Nov 4 13:02:57 EST 2018



> On Nov 4, 2018, at 4:54 PM, Gordan Bobic <gordan.bobic at gmail.com> wrote:
> 
> 
> 
>> On Sun, 4 Nov 2018, 16:48 Richard Elling <richard.elling at richardelling.com wrote:
>> 
>>> Note that random write I/O performance will be 120 IOPS per vdev (480 in first case, 240 in the other). It will be better for reads at approximately 720 and 300 respectively. Do some testing and maths to see whether that will suffice for your workload.
>> 
>> You’ve got this backwards: raidz does well on random writes.
> 
> 
> It always writes a full stripe - so it only gets 1 disk's worth of IOPS per vdev. Regular parity raid typically get half of that because it requires RMW for partial writes. RAIDZn does away with that by always writing a full stripe.

for raidz, “full stripe” doesn’t mean “all disks.” But more importantly, writes are most commonly allocated sequentially and aggregated. You’ll only really see small random writes on very full and very fragmented pools.

> 
>> It is the small, random read performance that suffers.
> 
> 
> If no recovery is required, the "parity" disks are free to service the next I/O required. So you get that part back in terms of IOPS.
> 

yes, the equation is a ratio of data to allocated size.


  -- richard
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://list.zfsonlinux.org/pipermail/zfs-discuss/attachments/20181104/afc6a229/attachment.html>


More information about the zfs-discuss mailing list