Btrfs RAID-10 performance

Linda A. Walsh dovecot at tlinx.org
Tue Sep 15 11:22:05 EEST 2020


On 2020/09/10 07:40, Miloslav Hůla wrote:
> I cannot verify it, but I think that even JBOD is propagated as a 
> virtual device. If you create JBOD from 3 different disks, low level 
> parameters may differ.
>   
----
    JBOD allows each disk to be seen by the OS, as is.  You wouldn't
create JBOD disk from 3 different disks -- JBOD would give you 3 separate
JBOD disks for the 3 separate disks.

    So for your 16  disks, you are using 1 long RAID0?  You realize
1 disk goes out, the entire array needs to be reconstructed.  Also
all of your spindles can be tied up by long read/writes -- optimal speed
would come from a read 16 stripes wide spread over the 16 disks.

    What would be better, IMO, is going with a RAID-10 like your subject
says, using 8-pairs of mirrors and strip those.  Set your stripe unit
for 64K to allow the disks to operate independently.  You don't want
a long 16-disk stripe, as that's far from optimal for your mailbox load.
What you want is the ability to have multiple I/O ops going at the same
time -- independently.  I think as it stands now, you are far more likely
to get contention as different mailboxes are accessed with contention
happening within the span, vs. letting each 2 disk mirror potentially doing
a different task -- which would likely have the effect of raising your
I/O ops/s. 

    Running raid10 on top of raid0 seems really wasteful




More information about the dovecot mailing list