a13antichrist wrote:qpio wrote:(4) when considering RAID-10 In the numbers you can see that RAID-10 only gives a slight advantage when running on the Atom CPU. Also, note that only write speeds where compared. The striping advantage of RAID-10 is only effective when reading.
This is misleading. Well actually, not misleading, it's just incorrect. The striping advantage comes into play for writing. It's the mirroring and the reading from both mirrors simultaneously that makes the [primary] difference for read results. XOR 'penalty' is of course also relevant, but less so with more powerful processors, as we've seen.
Incorrect? Oh wait...
Thanks for correcting my incorrect use of the word 'striping' there. If you are writing 100 GB to a 10 disk RAID-10 setup, you will be writing 20 GB to each disk. When reading you only have to read back 10 GB from each disk. So the advantage when reading is the mirroring indeed, which means you only have to read back chunks of data from each disk in a cough cough striped fashion.
With RAID-5 and RAID-6 these numbers are different (those are left as an exercise to the reader..) but there will be more heavy work for the CPU (XOR work) as well. You seem to imply that XOR also plays a role with RAID-10 but I fail to see how. RAID-10 can be efficiently implemented with a set of DMA IO commands. The RAID-5 and RAID-6 setups create more load on the CPU, but have effectively less IO to the disks while retaining the same level of data security.
So in the era of faster CPU's and while considering a 8-bay QNAP, I see no practical use of RAID-10 except for in a few really edge cases. Your last remark ".. but less so with more powerful processors" is exactly the point I was trying to make.
a13antichrist wrote:So it should be very clear to see that the "RAID-10 performance myth" is anything -but- a myth. You just have to have your context right.
Did you mean to say "nothing" but a myth? That would be a bit strong, but close to where I am currently on this topic.
I have seen RAID-10 setups with 40-50 disks in a single array. After splitting them up into 3 separate RAID-6 setups (with hot spares) these systems have produced better everything: more reliable, more throughput, more total storage, faster response times, less impact of one application's IO on another application's IO.
Obviously there are still cases where RAID-10 shines. For instance, in very large scale data storage, where data is distributed redundantly over multiple nodes with software like glusterfs (FS-level) or ceph (LUN-level). On the node level, RAID-10 might just be the best fit.
I was focussing on typical QNAP home scenario's where RAID-10 shines, and I can imagine none. By all means, let anyone share their experiences so we can all learn. I think the most valuable contribution would be to show some experimental results, and then trying to understand the underlying mechanisms.