On 07/03/2022 14:19, Natanji wrote:
Hello David,
what you propose here should work, but be aware that you are essentially
first writing all data to the 2 smaller drives and then reading from
them only when the mirror 1 TB disk gets added. If there are bad sectors
that only gets dicovered while reading (not writing), you might lose
data. A much cleaner way would be to make a backup of your data disk and
then first crate the array with all disks, then write the data back from
your backup once the disks are in sync. You should definitely make a
backup or you could potentially lose data when creating this array.
But then David is in an equal pickle - what happens if the backup
contains errors that only get discovered when reading, not writing ...
no, David's plan is just the same as yours, his "half a mirror" is your
backup, they suffer the same advantages and disadvantages.
I don't see the benefit of striping in your setup, by the way. For the
two smaller disks it might seem that this is beneficial for performance,
but since every write also needs to be made on the mirror, this will
essentially lead to your read/write head moving back and forth a lot,
no? Because if you write four consecutive chunks C1-C4 on the striped
side, two of these would be written to each smaller disk. And on the 1
TB disk side this would mean writing the sectors with only one
read/write head, each around 500GB apart from each other, and in a back
and worth fashion (although that might be optimized by queuing). So your
write performance could actually be *worse* than on a single drive, and
definitely not gain the 2x speed benefit of striping. For read
performance you might be okay, but nevertheless put unneccessary
mechanical wear on your 1 TB drive.
[C1 C3 ....] 500GB-1 [C2 C4 ....] 500GB-2
[C1 C3 ... (500G space) C2 C4 ....] 1TB
At least that's my understanding, someone correct me if I'm wrong and
mdadm handles such setups more intelligently (i.e. by laying out the
chunks completely linearly on the 1TB disk instead of 1-to-1 "mirroring"
the chunks on the 2x500G RAID0).
mdadm doesn't have a clue that the second device in your mirror is a
raid0. As far as it is concerned you have two completely SEPARATE raids,
a raid0-stripe, and a raid1-mirror.
So when reading from the mirror it should preferentially read from the
raid0 because it's the faster device. Which means it WON'T put
unnecessary mechanical wear on the 1TB device because it preferentially
will only use it for writing. And because the layer above sees it as a
1TB device, it won't be scattering writes all over the place. All three
drives would see a "stream write to the mirror" as a stream write to the
drive. A1,B1,A2,C1,A3,B2,A4,C2, ...
Cheers,
Wol
