Soft raid 6 or soft raid 10?

Question

linux

There is a server. There is no raid controller in it. The question arose, which is better (4 disks):

1) raid 10 - do not care about the loss of capacity twice, it worries that if two adjacent disks fail, the system will screw up, but oh, how you don’t want to. Now I raised 10, but there are concerns… Yes, and Ts.P. loads great;

2) raid 6 - any two disks can fail. However, there will be a loss in disk system performance, how critical is this? Does the CPU load the same as raid 10?

Z.Y. I'm talking about mdadm. The server will be used under middleware, for about a hundred subscribers with a possible increase in load to several thousand.

I also think how best to break the disk subsystem. In theory, almost all the free space should be allocated for the B.D. Of course, you can use lvm, but this is an extra load on the CPU again ...

In general, I would not want to screw up at this stage, and then redo it, because there is no second server for backup, and the service is important.

Answer 1

Raid6 "eats" more processor and it is not advisable to use it on 4 disks - the same two disks are lost with lower performance.
Stop at Raid10.
Although it's strange... if you have a critical service, then why can't you either buy more disks and/or increase the performance of the server as a whole.
And yes.
Backups for any need to do and do them on a third-party server / storage.

Answer 2

However, there will be a loss in disk system performance

It depends on the number of random write operations in small blocks. On average, it is considered that on a typical load like 65% Read 35% Write 100% Random, the performance of RAID-6 is approximately -40% of RAID-10 on the same number of disks.

Does the CPU load the same as raid 10?

Of course, it loads more, since Parity RAID calculates the checksum, and Mirror RAID simply writes the block twice by the controller, and does not consider anything by the processor. However, on modern powerful processors, this load is low.
However, everything again depends on the nature of the workload, on the number of small random records in the workflow.

Answer 3

Z would recommend 10. Performance is (much) better with mdadm, especially with sequential and completely random reads. The load on the CPU and on the bus is less. And most importantly: the search time is significantly lower.
About reliability. 6 is more reliable, but the resource of screws on 4 disks is developed much more strongly.
I would also recommend ZFS - in one bottle both mdadm and lvm and the file system itself. With more memory it works great.

Answer 4

Simply, any two disks can fail in raid 6, but not in raid 10. In general, I still tend to 6. For middleware, speed is not so important, is it?
Unfortunately no ZFS.