1

This is really a question of best practice.

In a datacentre environment whereby we have a single MD3200 DAS, and for the sake of the question, 6 expansion cages (MD1200s) how would you configure the RAIDS:

  • One single RAID per cage - i.e. a RAID10 consisting of all the physical disks in one cage, mapped as a single LUN to your servers/clusters. So essentially, we would end up with 7 RAID 10s. This way losing a cage will not affect your other RAIDs/cages.

  • Essentially the same, whereby we end up with 7 RAID10's but the logical disks are actually split across the cages. This way losing a cage may not result in the loss of any data, but could result in the loss of multiple RAIDs.

enter image description here

1
  • Are you using SAS disks? If so, you should be planning a multipath/ring topology from the controller.
    – ewwhite
    Aug 6, 2015 at 11:35

2 Answers 2

4

You shouldn't daisy-chain the enclosures if you can avoid it

That's also a lot of storage capacity to be planning. Can you give some more detail on what you're doing?

Are you using SAS disks? That's going to be important if you want a resilient cascaded JBOD setup. You'll need the dual-ported disks to make this work well.

See the graphic below. The SAS connections are in a ring topology so that the loss of a chassis can be handled without bring all other JBODs offline.

I suspect the Dell has guidelines on disk group and RAID membership... They are:

Support for RAID levels 0, 1, 10, 5, 6
Up to 120 physical disks per group in RAID 0, 1, 10 
Up to 30 physical disks per group in RAID 5, 6
Up to 512 virtual disks

So this may be a case where you structure your RAID groups to span enclosures. You have a lot of options, though. And complete JBOD or cable failure is uncommon.

enter image description here

2
  • ewwhite - each enclosure will contain 48TB. Essentially we will have a 4-node ESX cluster. It will be based in a datacenter on a MPLS environment servicing roughly 3000 users.
    – PnP
    Aug 8, 2015 at 10:24
  • I strongly 2nd this answer. Though a typical daisy-chain setup is the "default" configuration documented and recommended by Dell in most setup guides, the ring topology graph shown by @ewwhite is more fault-tolerant and capable of providing a slight performance improvement (it's also more in line with industry standard best practices, and IS documented by Dell somewhere as an alternate configuration option).
    – JimNim
    Aug 9, 2015 at 5:07
2

I don't know the MD3200 but I do know the HP equivalents very well - a lot of this depends on the cabling. Is each shelf separately cabled to the controller/MD3200 or are they 'daisy-chained' in some way?

The reason I ask is that if they're separately cabled then you have more options - it they're 'daisy-chained' then you need to consider the impact of losing the port or shelves closer to the controllers.

If they're separately cabled I'd be tempted to add an eighth shelf and configure your R10's 'vertically' - i.e. multiple 8-disk R1's, each using slot 1, the next using slot 2 etc. This way you could lose a whole port/cable/shelf/disk and have zero impact on your ability to read/write - potentially up to half of your shelves in fact (if it were the 'right' ones anyway). By the way if you were using R6/60 you could do the same with your seven shelves but for a R1/10 obviously you'd need an even number of shelves.

If they're 'daisy-chained' then I'd be tempted to look at each side of the R1's that make up your R10 to ensure that they're split across these 'chain-blocks'. So for instance if shelves 1 and 2 are 'daisy-chained' together and so are 3 and 4 then I'd try to create each of my R1's so that half is on 1 or 2 and the other half is on 3 or 4. This way you could lose the port, the cables or the shelf pair again without impacting read/write-ability.

5
  • They are indeed daisy chained.
    – PnP
    Aug 6, 2015 at 11:09
  • Do you know how in more detail? If so I'll update my question based on them if needed.
    – Chopper3
    Aug 6, 2015 at 11:11
  • Yes, let me add a small diagram to OP.
    – PnP
    Aug 6, 2015 at 11:12
  • Updated. Green are daisy chain links to each expansion. Blue area are the expansions. Top unit is the controller which is linked to cluster hosts.
    – PnP
    Aug 6, 2015 at 11:16
  • Ahhh - I see, so they're all off one set of controller ports - in truth that would bother me as if you lost those ports you're dead in the water - but given what you've said I'd suggest you use the internal disks inside the controller to create a single R10 and store your most 'gotta stay up' data on those as they wouldn't be affected by the loss of those external shelves. The I'd do exactly as you suggest with the remaining 3 shelves, one per shelf - I think that'd be the best you could do.
    – Chopper3
    Aug 6, 2015 at 11:21

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .