Configuration for Open Source Operating systems with the SAM-SD Approach
-
I wasn't sure how far you would really need to take for mission critical data. I was assuming that at some point you only want the data mirrored once. Whether that was mirrored at the drive level or at the system level I wasn't sure. I'm assuming if you have huge storage arrays you'd want to have two SANs with JBOD Arrays, mirrored across each other? or would you want to have a RAID6 setup in each SAN and then each SAN mirrored to each other? this way (like we talked about earlier) you can have a few drives go bad and not have to rebuild the entire array. In my mind the second way (with RAID 6) seems to make the most sense?
-
@GotiTServicesInc said:
I wasn't sure how far you would really need to take for mission critical data. I was assuming that at some point you only want the data mirrored once. Whether that was mirrored at the drive level or at the system level I wasn't sure. I'm assuming if you have huge storage arrays you'd want to have two SANs with JBOD Arrays, mirrored across each other? or would you want to have a RAID6 setup in each SAN and then each SAN mirrored to each other? this way (like we talked about earlier) you can have a few drives go bad and not have to rebuild the entire array. In my mind the second way (with RAID 6) seems to make the most sense?
No large, mission critical storage system is using RAID 0 per node. If you think about drive failure risks and that the time to rebuild a node over the network is long the risk would be insane. Say you have two 100TB SANs. If you lost a single drive on one node, you've lost 100TB of data and all redundancy. Poof, gone, one drive having failed.
Now you failover to your second node. You now have 100TB on a SAN on RAID 0, zero redundancy!! Assuming you could replace the failed SAN in four hours and start a RAID 1 rebuild across the network and assuming that you have 8Gb/s Fibre Channel (which is 1GB/s) that would take 27.7 hours to copy back over assuming zero overhead from the disks, zero overhead from the network protocol, zero delays and absolutely no one accessing the SAN in any way during that 27.7 hour period. Any reads will slow down the network, any writes will slow down the network and require additional data be transferred.
So a theoretical best case scenario restore is 32 hours during which you have 100TB sitting on RAID 0 without anything to protect you from the slightest drive blip. And suddenly you would be vulnerable to UREs too, which would be expected to happen rather often in an array that large.
More realistically you would expect to be restoring for at least tree days if you had dedicated 8Gb/s FC and more like a week and a half if you were only on dedicated GigE iSCSI.
That's a long time to pray a RAID 0 doesn't have an issue.
-
And that rebuild scenario would come up pretty often because drives fail often when you have big numbers of them. This isn't a theoretical edge case, you would expect this to be happening a couple or a few times a year on a really large SAN cluster like this. That's 200TB of drives after redundancy. Even with big 4TB drives, that's 50 of them. They are going to fail from time to time.
-
so for a large setup like that (50 drives), would you want to do a raid 6 per 10 drives and software raid 0 them together to allow for a quicker rebuild time with more drives being able to fail simultaneously?
-
With a100TB array, even RAID 6 is getting to be a bit risky. A rebuild could take an extremely long time. A RAID 6(1) is really the riskiest thing that I would consider for something that is pretty important and RAID 10 would remain the choice for ultra mission critical.
-
@GotiTServicesInc said:
so for a large setup like that (50 drives), would you want to do a raid 6 per 10 drives and software raid 0 them together to allow for a quicker rebuild time with more drives being able to fail simultaneously?
You never mix hardware and software RAID, not in the real world. That's purely a theoretical thing. Of course you can, but no enterprise would do this. Use all one or all the other.
-
@GotiTServicesInc said:
so for a large setup like that (50 drives), would you want to do a raid 6 per 10 drives and software raid 0 them together to allow for a quicker rebuild time with more drives being able to fail simultaneously?
You would either accept the risks of RAID 6, go for full RAID 10 or blend the two with RAID 60. At 25 drives you are just large enough to consider RAID 60 in a practical sense.
http://www.smbitjournal.com/2012/11/choosing-a-raid-level-by-drive-count/
-
Choices with 25 drives per node with RAID 60 would be...
5 sets of 5 drives. That's it. Only one configuration.
So the RAID 60 overhead would be 15 usable drives and 10 parity drives. Not a good mix at all. You lose a ton of capacity as well as performance for only moderate additional protection.
-
Of course to get to our capacity of 100TB we have to consider that assuming 4TB drives....
RAID 0: 25 drives
RAID 6: 27 drives
RAID 60/5: 35 drives
RAID 10: 50 drives -
So really the only correct solution is a RAID 10 setup with mirroring over FC or iSCSI? I still feel like you'd be stuck with a whole lot of rebuild time if a drive failed in one of the arrays, although the rebuild time should be faster as long as you don't have a failure in both JBOD arrays on a single server?
-
But in a 20 drive array, you have 10 RAID 1 drives that are RAIDed 0 over them. So a single drive failure only requires the mirroring of a single drive.
-
If I understand all this right, I think that RAID 10 would have the fastest rebuild time since it only has to rebuild a single disk from its partner in the same server. It doesn't have to fight network congestion or anything like that quite so much.
-
@GotiTServicesInc said:
So really the only correct solution is a RAID 10 setup with mirroring over FC or iSCSI?
You can't really mirror over a SAN block protocol like that. In theory you can, but it is incredibly impractical and nothing really leverages those technologies for that. You mirror either at a RAID level or via a replication protocol like DRBD or HAST.
-
@GotiTServicesInc said:
....although the rebuild time should be faster as long as you don't have a failure in both JBOD arrays on a single server?
There is no JBOD. You wouldn't have that in any situation.
-
What is it that folks like HP, NetApp, EMC, et al. do? Do they do it at the block level or via some type of other method like DRBD?
-
@GotiTServicesInc said:
So really the only correct solution is a RAID 10 ....? I still feel like you'd be stuck with a whole lot of rebuild time if a drive failed in one of the arrays
A drive failure on RAID 10 is always the same no matter how big the array is. A drive resilver on RAID 10 is always, without exception, just a RAID 1 pair doing a block by block copy from one drive to its pair. That's it. So if you were using 4TB drives like our example, the rebuild time is the time that it takes to copy one drive to another directly. That's all. It's the smallest rebuild time possible for any RAID system. You really can't make it faster than that.
-
@dafyre said:
What is it that folks like HP, NetApp, EMC, et al. do? Do they do it at the block level or via some type of other method like DRBD?
Always something like DRBD. Normally it is proprietary, but not always. NetApp does not have a stable, working replication last I heard. Some vendors actually just use DRBD (Synology, for example.) Some make their own replication products. But the general theory is the same - write to both at the same time, read from the local.
-
@Dashrender said:
But in a 20 drive array, you have 10 RAID 1 drives that are RAIDed 0 over them. So a single drive failure only requires the mirroring of a single drive.
Correct. Which doesn't mean that recovery is seconds or anything like that. But we measure the recovery time in hours with minimal impact rather than days or weeks with massive impact.
-
If you move a RAID 10 array from 4TB drives to 2TB drives you essentially cut drive recovering time in half, too. So you can balance things depending on your needs.
-
Handy thing to think about.....
Parity RAID: Drive resilver time is determined by a combination of drive size and array size.
Mirrored RAID: Drive resilver time is determined only by drive size.
