On 19/5/21 09:48, antlists wrote:
On 18/05/2021 17:05, Phillip Susi wrote:
Wols Lists writes:
When rebuilding a mirror (of any sort), one block written requires ONE
block read. When rebuilding a parity array, one block written requires
one STRIPE read.
Again, we're in agreement here. What you keep ignoring is the fact that
both of these take the same amount of time, provided that you are IO
bound.
And if you've got spinning rust, that's unlikely to be true. I can't
speak for SATA, but on PATA I've personally experienced the exact
opposite. Doubling the load on the interface absolutely DEMOLISHED
throughput, turning what should have been a five-minute job into a
several-hours job.
And if you've got many drives in your stripe, who's to say that won't
overwhelm the i/o bandwidth. Your reads could be 50% or less of full
speed, because there isn't the back end capacity to pass them on.
Cheers,
Wol
Jumping into this one late, but I thought the main risk was related to
the fact that for every read there is a chance the device will fail to
read the data successfully, and so the more data you need to read in
order to restore redundancy, the greater the risk of not being able to
regain redundancy.
So, assuming all drives are of equal capacity, you will need to read
less data to recover a RAID10 than you would to recover a RAID5/6, thus
a RAID10 has a better chance to recover.
1) speed of recovery (quicker to read 1 x drive capacity instead of n x
drive capacity even if all in parallel) (unless you can sustain full
read speed on all devices concurrently I guess).
2) equal load on the single "mirror" device from RAID10 compared to load
on ALL devices in the RAID5/6.
3) lower impact to operational status (ie, real work load can continue
without impact on all reads/writes not involving the small part of the
array being recovered, equal impact for read/write that does involve
this part of the array).
Right?
Regards,
Adam
