I’ve had the usual-suspect vendors self-report 20MB/s or thereabouts RAID rebuild rates to me. Isilon reports, for a large cluster, that the RAID rebuild mechanism is affected in a distributed fashion by every head in the system. Rebuild rates are 400MB/s and beyond. That’s 20X conservatively. For a full sized cluster, actual advertised rates are far higher than 400MB/s. And I believe them.

As for your random workload comment, I am sure you are right. Equally, I don’t particularly care. Storage is a tool, and for the proper use case, Isilon’s product is good one.

Lastly: my remark was not “unfounded”. You may believe it is mistaken, but unfounded it is not.

Joe.

I remember that for a long time, they were publically very negative about the general block-based SAN market until they were in a position to deliver the technology to support it.

Perhaps we should take a look at their record with Raid 5, which may have conditioned this position. Why did they go for so long with Raid 4 backends when it is unsuitable for IO intensive applications? Did they ever, during that time, acknowledge the benefits of Raid 5 ?

Raid DP is a diagonally striped, dual Raid 5 scheme … just simpler to compute than Raid 6. As I recall, similar algorithms were first researched and published by IBM. It would be interesting to find out if Raid DP developed in-house or purchased from an outside source and then patented ?

I may be wrong, but it seems that the actual RAID 6 controller was first implemented by a small company based in Taiwan, closely followed by a lot of publicity by Intel for their R5/6 hardware-assisted Xscale processors. Open source example of Raid 6 has been around for a while.

No, I am not in marketing.

Maybe because people actually want performance for a dataset other than streaming media?

That “AWESOME” protection isilon provides falls over and dies as soon as random small file workloads (aka: REAL WORLD workloads) are thrown at it.

Your 20x number is completely unfounded.

I think the RAID-6 mantra is a bit overused. It’s protection from data loss that we need, and for most architectures that means minimizing exposure time to the the last allowable failed component… not some specific RAID level.

Joe.

I’m sitting next to one of the first boxes they ever produced and it’s RAID4, so I still don’t know what you’re talking about.

Furthermore, if they were advocating RAID1 as you claim, it should be trivial for you to provide us a whitepaper, or document of some sort verifying those facts. I can readily find them for both RAID4 and RAID-DP.

Again, why would I put a second box in front of block storage and add more management overhead when it buys me nothing? Spitting out a bunch of industry coined phrases doesn’t make it a better solution. You must be in marketing jamming all that into one paragraph.

And finally, if NetApp weren’t the first major storage vendor to adopt raid-dp, as you were so quick to brush off, who was? You seemed to have left that tidbit out.

Parity-based RAID is beginning to breakdown, RAID-5 in the terabyte+ disk world is a non-starter; RAID-6 in the 2 terabyte+ disk world is a little scarey don’t you think? So, some kind of mirroring needs to happen! It all comes down to how you do this, how you minimise rebuild-times, maximise performance.

I actually think it is funny that Moshe (who never really got RAID-5 etc) actually pops up at a time when mirroring may come back to the fore. And the irony of EMC pointing fingers at a fully mirrored box considering the crap they used to spew about RAID-5 before they had a workable implementation is absolutely fantastic!!

The XIV array does have a number of weaknesses/issues but I don’t think the fully mirrored situation is one of them. Now non-disruptive code upgrades, 1 GBE backplane, relatively limited capacity and the like of async replication is alot more of an issue.

The system does have some interesting properties. Virtual “RAID 10″ as a distributed system is kind of cool, but what would be cooler is something that doesn’t eat the bytes and drive the “per usable GB” off the scale.

The product’s choice of “RAID 10″ is defended by Moshe on the grounds that advice in industry is to “mirror everything”. This is fine for what its worth, but me, when I mirror my systems, I like the mirrors to not be on the same system, if you know what I’m saying. Ever had it rain in your server room? (I have!).

I confess my assessment of this product could be myopic: I am very very aware of Tier-2 systems, but tend to be less well informed regarding true first class Tier-1 systems. The target market for this product appears to be folks who want, but cannot afford, Tier-1 SANS.

Joe.

I think you will find that earlier NetApp appliances used general -purpose motherboards and software RAID1 …. later followed by software RAID4. RAID DP is a recent addition…. no, they were not first to demonstrate dual parity.

What exactly are you talking about?

First, why would I put a fileserver in FRONT of block based storage. Blue Arc and NetApp integrate it, and save you a lot of unnecessary hardware, expense, and management overhead.

EMC and NetApp defending RAID1? What? NetApp has *NEVER* used Raid1. EVER. They used RAID4 from the start, and moved to RAID-DP after that (And I believe were the first to use dual parity).

The REASON they moved to RAID-DP was the integration of larger and larger ATA drives which are more likely to fail, and have longer rebuild times, thus requiring *better* protection.

Having to triplicate disks with single point failure backend controllers to drive these disks … is not really a great innovation in today’s data center power & uptime climate. Also, we have the complexity of applying multiple backend ‘controller’ resources to move data over a 1Gbit switch to deliver fast rebuilds…. and ignoring disk drive write bandwidth. In addition, there is a ‘single-point failure’ backend … single controllers driving 15 disks. These controllers that can’t be hot-swapped … this is what you get for using ‘commodity’ server boards as controllers… so there is the need to protect the system with complete, multiple ‘spare’ storage enclosures … which, in turn, requires multiple disk rebuilds…. and more power.

All of this is not very well thought out…. I think it started with the inability to provide fast RAID 6 on commodity hardware and the rest of architectural ‘innovation’ followed from there.

I remember EMC and NetApp defending (for years) the benefits of RAID1 until they ‘discovered’ the need for RAID5 …. plus some very recent statements by EMC personnel that RAID 6 was not required … and now it is a large ‘benefit’.

It is interesting to note … how so many… so well educated people …are suddenly willing to change their tune. I suggest that we can expect the same from IBM… they are now probably retrofitting the system to a 10G fabric… trying to figure out how to make their ‘commodity’ backend controllers hot-swap… perhaps RAID 6 protection within the disk enclosure … this should be interesting.

Please don’t send the cube.