Re: Disk I/O slowness

Michael Stone <[email protected]> Wed, 5 Aug 2026 12:28:05 -0400
Newsgroups	gmane.linux.debian.user
Message-ID	<[email protected]>
On Wed, Aug 05, 2026 at 12:41:36PM +0500, Alexander V. Makartsev wrote:
>But if these two failed drives will be in the same mirror group then 
>the whole array will be lost.

Yes, and the drive that gets stressed the most in that critical 
situation is the *one* that's being rebuilt from.

>>And rebuilds can take a *long* time (like, days).
>You probably thinking about RAID-0+1 array configuration, which is not 
>RAID-10.

nope; see below

>Suppose we have 8x4TB drives. In case of a single drive failure the 
>RAID-0+1 array would have to rebuild the whole 16TB worth of data (one 
>mirror group of 4 drives) which indeed could take a long time on a 
>busy server,
>but with the RAID-10 array it will rebuild one mirror group worth of 
>data (4TB) which is the size of single drive.

The rebuild time is dominated by the write speed of the target drive. 
4TB were common in large arrays 10 years ago. Now 20TB is fairly common.
Transfer rate on that is less than 300MB/s. At 300MB/s that 20TB takes 
18h to write. Realistic performance is probably more like 150MB/s, or 
36h. Extrapolate for 36TB drives. If the array isn't idle, it will take 
longer. (And few large arrays are purchased to be idle.)

>RAID-5 and RAID-6 rebuild of 8 drive array would take a very long 
>time, because not only it will have to recover (not just copy) the 
>data from error correction data spread across all drives

Reading the data from all the drives is a positive rather than a 
negative: it means the controller is reading from say, 7 drives to build 
data on 1. That means a lot more idle cycles on the 7 as opposed to 
having 1 drive 100% busy (which matters if you're trying to use the 
array during the rebuild).

>it also will 
>have to rebuild and recalculate the error correction data for the 
>failed drive for the entire array.

That computation is trivial on any modern hardware--the only bottleneck 
is drive speed.

>>You can get double disk failure tolerance on RAID-10, but then 
>>you're down to 33% usable space.
>>
>This statement looks false to me, can you elaborate?
>AFAIK the RAID-10 array configuration, build from even number of 
>devices (4 drives or more), will always give at least 2 drive fault 
>tolerance, except in the situation I've described above, and the 
>usable space for the 4 drives RAID-10 array always will be 50%.

"fault tolerance, except" isn't fault tolerance; if you want to tolerate 
the failure of *any* 2 drives in RAID-1 you need 3 drives. The up side 
to that is that you can then identify silent errors via voting.
signature.asc (application/pgp-signature, 833 B)