Re: Disk I/O slowness

Hi,

On Fri, Jul 31, 2026 at 12:41:24PM +0500, Alexander V. Makartsev wrote:
> On 7/31/26 01:09, Andy Smith wrote:
> > On Linux mdadm you can also do a RAID-10 across three drives which still
> > allows for the loss of any single drive and provides both better read
> > and write performance compared to RAID-5 of the same devices. The
> > capacity would be 1.5 drives as opposed to 2.
> That is an interesting idea.
> The RAID-10 configuration you proposing should then consist of 3 RAID-1
> (mirror) groups, so 3 drives must be partitioned in halves and to finish
> RAID-10 array, there should be a RAID-0 (stripe) array created across all 3
> groups.

No, Linux mdadm just does it natively. Just as you do not need to make
sets of RAID-1 and then put RAID-0 on top for RAID-10 of even numbers of
drives, so you do not need to do any partitioning tricks with odd
numbers of drives. mdadm just directly works it out.

The biggest downside is that mdadm RAID-10 is hard to alter the shape of
if you add or remove drives later.

This is all pretty old tech by now of course. Copy-on-write filesystems
like btrfs and zfs do things in a much better way. I had high hopes for
bcachefs, alas…

> Will the RAID-0 (stripping) layer of RAID-10 configuration significantly
> improve the write performance of the array in comparison to a basic RAID-1
> array?

It's not going to be dramatic when you compare a 2 drive RAID-1 to a 3
drive RAID-10  to a 3 drive RAID-5, but yes because stripes go across
multiple drives whereas with RAID-1 there is only ever one drive worth
of write IOPS.

Thanks,
Andy