Storagezilla

The ARM server apocalypse

I noticed earlier tonight that Taiwanese vendor MiTAC will sell you an eight core 64 Bit ARM server.

Which, proving how flexible the ARM licensing model is, happens to be months ahead of delivery compared to ARM’s own 64-Bit Cortex-A50 and NVIDIA’s always on going 64-Bit ARM CPU development efforts.

Linux already runs on ARM and while Microsoft continue to play it cagey on their thoughts about ARM binaries for Windows Server, owning a Microsoft Surface I can tell you that Windows does run on ARM today even if there are still massive performance optimisations being pushed out on a regular basis. Windows RT is so much Windows that patch Tuesday requires the multiple reboots for Windows Update installation that Windows 8 on x86 does.

There’s no embedded OS monkey business going on with Windows RT, it’s Windows straight from the source and with that comes all the awe & terror of living with Windows.

While Windows RT is Windows 8 on ARM in 32 Bit Mode it proves ARM like PowerPC or DEC ALPHA in the Windows NT 4 days, is a viable target platform technically. Even if the business case for Windows Server just isn’t there yet.

That business case will come more from serving static content to people over the web than it will be from running computationally heavy distributed applications, but that’s a major slice of the server market today as it’s the presentation layer for all this ‘cloud’…stuff. –cough-

MiTAC are using a SOC (System On Chip) from Applied Micro Circuits Corp, the same supplier HP signed up for it’s yet to appear ARM powered Moonshot systems but with AMD promising a 2014 debut for it’s ARM server processor I can image the major server vendors are already sniffing around the major ARM CPU manufacturers to see how close they can get to x86 server performance and how much ‘good enough’ processors will cost them.

With AMD telling people it costs them $400M and takes three years of development to crank out an x86 server processor design while on the other hand it costs them somewhere to the order of $30M and 18 months of their time to crank out an ARM server processor design it’s all going to come down to how many generations of development the ARM licensees are going to need to get to good enough performance.

Intel loom large as they do in every processor conversation and are aggressively positioning Atom to meet the threat head on. They have an incredible advantage in the process area where their absolute control from design to manufacture nets them performance improvements as part of manufacturing improvements but their investment in process has to always be massive as their competitors are operating without the spend on manufacturing process development. The ARM collective simply show up to Samsung or TWSC, or TI or Global Foundries, ask for a zillion processors to run off the manufacturing lines and wonder if they can have them by Tuesday? 

If, by 2015, 64-Bit ARM makes it within striking distance of x86 server processor performance, the sheer number of different suppliers will drive prices down well below what Intel can match and before the end of the decade the ARM collective could go from zero to 15 or 20% server CPU market share.

When you consider how desktop computer CPU manufacturer Intel, who was growing like crazy during the golden age of Wintel expansion, demolished the higher performing mega-margin server architectures of the 90’s, the DEC ALPHAs, PA/RISCs and SPARCs of the world as it made IBM POWER (Life assurance provided by the IBM Mainframe monopoly) a niche within a niche, this is one of those industry inflection points which could hammer Intel very hard.

The way Intel’s expansion into the server market was funded by tens of millions of PCs landing on peoples desks and in their homes, ARM’s expansion is being funded by the hundreds of millions of smartphones in people’s pockets and tablets landing in their hands.

ViPR, where automation meets intelligence.

Having just been asked the question at 3:00AM this morning (Sleep is for the weak) I thought it was time to define ViPR as I understand it.

As *I* understand it mind you. It’s creators, EMC sales & marketing and who knows who else might have a different opinion than I do, but with that said to me:  

ViPR is a policy based configuration, management, monitoring and self service provisioning for heterogeneous storage infrastructure. It also provides additional data services to the storage in that infrastructure.

Begin the complaining emails, dial up the snarky comments, dispatch the marketing killbots, but brass tacks that’s what ViPR is today.

It’s not about the encapsulation of LUNs behind an array controller, it is about logging into a portal, filling out some fields & checking some boxes to specify performance, availability requirements and so on, then hitting a button for ViPR to make everything you just specified happen.

Zoning, masking, quality of service, LUN or File system creation, access control, snapshot policy, replication type and the destruction of all of those when you decide you’re done with them, performed by automation. And when all that is done ViPR gets out of the way. It monitors. meters and reports but your data doesn’t flow through it.

You no longer have this brand of array controller or that brand of array controller you have storage pools which can provide you this performance, that availability and these kinds of data services. Logical management abstractions of physical resources with physical connectivity.

What about the data services you don’t have? What if you want object storage but want to reuse your existing storage? What if that existing storage is feature poor? Well ViPR can use that storage as a target for data services ViPR itself will provide. ViPR can use your NAS storage as object storage target, provide Atmos, SWIFT and S3 API access and allow you to access those objects as files if you so choose.

Want to pour a lot of your existing data through Hadoop? No problem, ViPR can present that data from where it is as HDFS. No need to copy that data into a Hadoop file system and then run Hadoop against it.

Those are two of the additional data services ViPR can make available above and beyond what might have come in the storage array from your provider and it’s my understanding the ViPR team will be adding more over time.

Bottom line: Policy based automation for the stuff you have and the ability in software to add the stuff you don’t.

(Now I am going to bed, tomorrow today is a travel day.)

Data Domain. Oracle RMAN multiplexing now welcome.

Without writing a magnum opus on the topic one of the features I really like in the new DDOS 5.3 release for Data Domain systems is the ability to deduplicate multiplexed Oracle RMAN backup streams.

Multiplexing as a backup concept is a hold over from the days when you needed to keep tape drives writing data at a minimum throughput to prevent rewind operations from occurring. You did this by interleaving backups from different systems, queuing up the writes and sending them all to the same tape drive. Net result you kept the tape drive fed during write but the restore speed was significantly lower due to seek operations to read the interleaved fragments back. Disk eliminated the need for multiplexing.

Oracle DBAs, having to maintain overall system responsiveness, have a different set of issues and juggle channels, disk buffers and data rate transmission when it comes to moving as many database files to a backup target as quickly as they can. The result of this means as they balance the resources available to them it can lead to a multiplexing situation where multiple reads are combined into an interleaved write as that’s the best use of their available system I/O.

Multiplexing, like compression or encryption is awful for data deduplication but unlike either of those it isn’t down to re-encoding the information or eliminating commonality, it’s a direct result of the metadata inserted in amongst the files that have been interleaved together. The metadata causes the deduplication ratio to fall when the deduplication system fails to identify it as something apart from, not in, the database files themselves.

In DDOS 5.3 however the Data Domain operating system can identify all this metadata as being separate from the database files and therefore provide a backup deduplication ratio on par with a non-multiplexed backup of the same information.

And I think that’s pretty cool.

Busiest EMC World for Backup Recovery Systems *ever.*

Against the backdrop of the ViPR announcement and the predictable competitor sniping which followed..

We get it. You’ve done nothing to advance storage outside of your array controllers and as such now have to take the position that no one should. I’m betting as is usually the case there were teams established immediately after ViPR was announced to chase after EMC and we’ll all start reading the strategic leaks in six to nine months that you ‘have something coming’.  Too late, the boys & girls in the Seattle office are already setting the pace with code while you have yet to start drawing pictures in PowerPoint.

..this was probably the busiest EMC World for me in years.

It’s also indicative of how things are going for Backup Recovery Systems in general.

There was a time not so long ago where backup & recovery inside EMC had three demo pods at the EMC booth, less than a handful of sessions and the Thursday morning asscrack-o-clock birds of a feather session.

I’d show up, hang out at the booth and answer some of the weirder questions. Attend every other EMC product update session except my own as there wasn’t much going on with them at the time, then roll into the birds of a feather session and with any specific question would insult at least a quarter of the attendees with my answers.

Life was pretty easy. And dull as hell. 

Then BRS became this data protection juggernaut in the industry and that easy boring life vaporized. We have so much going on at the moment with the BRS hardware platforms and software suite as well as the shift going on in the data protection industry in general that what other people are doing inside EMC not associated with what I’m focused on becomes extra-curricular. Driven not by tedium (as it was when nothing much was happening with the anaemic product set years ago) but by curiosity and necessity.

And that’s great.

Instead of being the small cluster of ‘backup weirdos' inside EMC in the days before BRS, the question to the non-BRS parts of EMC is now ‘What’s you’re data protection strategy?’ If the answer that comes back isn’t satisfactory it becomes an Executive level issue.

BRS CTO Stephen Manley & BRS President Guy Churchward’s keynote is worth watching. I get to shout at people from in the audience, Stephen makes us do tricks and Guy gets to publically introduce himself as head of the largest Data Protection provider in the world.

And does so while tweaking EMC President David Goulden.

Numerous times.

Giveaway: $150 EMC World discounts.

When I have stuff to give away you’ll get it, in this case I’ve a $150 off your EMC World Registration to give away to the first three commenters.

Include a valid email address so I can contact you directly. If you’ve used an early bird or discount code already don’t bother as it’ll knock out the previous code, but if you’ve already registered without any early bird or discount codes it can be used after the fact to gain the discount.

XtremIO: XIOS 1.0 (Now shipping)

Since I don’t write anything ahead of time I’m always under the gun when it comes to product launches. You can look at pretty pictures and contemplate your navel elsewhere, this post will last as long as my morning cup of tea so here are just the facts. 

At time of writing the current shipping release of the XtremIO Operating System (XIOS) is version 1.05. In this release of code the management system, X-Brick N-way clustering via RDMA over InfiniBand, the global memory pool and the zero configuration XtremRAID (Not a real name, I use it to differentiate XtremIOs Flash protection scheme from the disk array stuff we’re used to) systems are fully operational allowing for XtremIO X-Brick scale out performance functionality, no LUN carving requirement and no requirement for performance tuning.

Inline data deduplication is enabled by default and is global inside the system, every X-Brick has access to a global deduplication cache meaning unique 4K blocks are written to flash only once and N-way access means all SSDs are accessible to all X-Bricks and therefore wear level at the same rate. Any volumes you make available out the front end can be made available via any port on any X-Brick you choose.

Connectivity is via 8Gb FC or 10Gb iSCSI and VAAI (Zero block, X-Copy. UNMAP/TRIM, ATS, whatever) is fully supported for vSphere 4.1 and vSphere 5.0 onwards. Snapshots are available but there will be a code upgrade adding more granular data services later.

The use cases for the system are the same as they ever were, VDI, Server Virtualization and Databases. So pretty much anything that requires a lot of small random IO.

The above is no substitute for reading the release notes, but in a small nutshell that’s the 1.0 release of code shipping today.

A quick note on Directed Availability.

DA is an EMC business process concept, not a technology concept.

Since 2011 (Maybe even further back) when EMC launches a net new product or a product where there’s a significant amount of new or changed functionality the product is first shipped to market through Directed Availability. New Data Domain code, new VMAX, new Isilon, new NetWorker, new Avamar  new whatever is sold first using the directed availability process.

The product is for sale, orderable and shipping but there are metrics which control how many systems are manufactured or how many copies of the code are shipped. Okay, so metrics such as what you might well ask?

-Number of support & delivery engineers trained on the product.

That’s pretty much self explanatory as if you have seven thousand people selling a product you want to make sure you don’t just have three people back in the office to support it while their co-workers sit in a classroom learning the new stuff. The more engineers you have ready to deploy and support the more products you build and sell.

-Cumulative production run time.

You set down how many hours of continuous production operation the system must meet out in the wild before you scale the manufacturing and delivery plans for the product to meet projected market demand.

-Supply chain quality & availability.

Lets say you’re sourcing new parts for a new offering, if you find out that every night when darkness falls one of those parts creeps out and eats kittens, you’ll want to know that before you’ve ten thousand systems sitting around the world. On the opposite side of that lets say you’re using a component currently only deliverable by one manufacturer, they say they can deliver the 50,000 of them a month you have believe you’ve demand for but they only deliver you 5 because they have problems manufacturing them. You want to find all of that out before you take $100M in orders from customers.

So those are examples of the metrics in the DA process. The more your business is ready to absorb everything that comes with selling a new technology the more of the technology you can plan on selling.

Getting away from the part that puts things in cardboard boxes or on download sites and back to the tech lets ask..

Whatever happened to Project Thunder?

Amongst talk of Xtrem cards, Xtrem code and XtremIO, Project Thunder is notable through it’s absence. But it’s in all of what they’re announcing today and are going to announce ‘later in the year’.

Let me explain.

You might recall that Project Thunder was what I/they called ‘server networked flash’. A scalable PCIe based solution networked together with InfiniBand to deliver millions of IOPS at a very low latency. So the team wrote the code, built the product, trotted it out for some behind closed door meetings with customers, did some proof of concepts and when they ran through all the customer feedback it boiled down to one thing.

Like it a lot, but it needs to be more like a storage array.

Indeed were you to go through all the feedback you’d find that XtremSW Cache (The artist previously known as VFCache) and XtremIO cover every use case Thunder was designed to address between them. So the Thunder hardware has been shelved and the Thunder data path and data services software has been absorbed by XtremIO, XtremSW Cache and parts of the XtremSF Suite ‘later this year’.

That’s pretty much all I’ve got for this post. No changes to XtremSF Cache besides the rename and ability to use the new XtremSF DAS cards in their various sizes, but you can expect a major version number increase ‘later this year’.

And that was the final swallow so we're done. See you in Vegas…I mean, ‘later’.

(And don't forget to backup your data regardless of if it's on Flash or not.)

Time to get Xtrem

Tuesday.

Virsto, how do I back it up?

The moment anyone inside EMC or VMware buys anything, ever, the first question I get from an end user or a sales person is ‘How do I back it up?’

VMware acquires Virsto on Monday evening and starting at 7:00AM on Tuesday morning I’ve been getting ‘How do we back up Virsto?’ questions.

Fair question but it plays to not understanding what Virsto.

When you look at Virsto for vSphere (There is a Hyper-V flavour where you run in the parent partition but regardless of how they’re implemented it’s turtles all the way down) you see serialisation of write I/O at a *physical* host level through the use of logging, the contents of those logs being asynchronously written to a shared pool of physical LUNs (Usually carved from an external storage array) probably abstracted away with a distributed file system presented to all physical hosts.

It’s most definitely more complicated than that as I only had the elevator pitch in a crowded corridor as I was off to find some ice cream in San Francisco but think of it this way, you’re running some of the stuff you find in the storage array controller today across your physical hosts.¹

When you understand at what level Virsto operates at it becomes clear you backup and restore your virtual machines the same way you always did. Using VADP with CBT & CBR or Guest level backup and recovery with products like VDP/VDP Advanced, Avamar, EMC NetWorker or other.

When it comes to backup and recovery the best thing VMware did was shoot VCB in the head, the worst thing they could probably do would be to break the VADP workflow entirely. Virsto does not break the existing workflow.

What could change is the longer term is additional usefulness of the snapshots taken during VADP operations. Lets face it VMFS snapshots are incredibly horrifying for a whole number of reasons not worth repeating here but Virsto can solve many VMFS snapshot limitations.

Whack in some VFCache cards to keep those Virsto physical host logs closer to the CPU and with random reads you have yourself a real party when it comes to test & dev snapshots, but I haven’t looked at that at all as it was a thought that occurred to me at the time so I’m certain your mileage will vary.

Bottom line, regardless of VMware now having their own storage controller layer for their software suite your backup and recovery workflow does not change. 

¹‘So VMware is competing against storage array vendors now?’ No I don’t think so, at least not any more than they already were. if you consider Virsto to be a write serialization technology it’ll improve the overall random write performance of any external storage array providing the underlying storage. In that sense it makes array vendors offerings more valuable since hosts can get more work done in the same amount of time as they could before.

On the other hand if your entire value prop is wrapped up in the fact that VMFS snapshots suck donkey ass and your array snapshots do not then this will be a net negative for you going forward.

Dell: Slash & burn.

While Michael Dell might be crowing like a cockerel at taking his company private the question for Dell customers should to be this…

Now carrying an extra $15 Billion in debt, what’s Dell going to slash & burn first?

With new creditors to be paid what’s Dell going to stop investing in?

You can put all this talk Mike is floating about new acquisitions aside, anything Dell might have been able to afford they could have bought as a public company and the bottom line is that Dell will owe it’s gang of four bankers $15B tomorrow which it did not owe them yesterday.

Not only are Dell out of the market for any major acquisitions but their R&D spending position is now questionable as it will no longer be reported.

While Mike may be ready to give himself to his corporate namesake body & soul, the bankers he’s borrowing from will be single minded in pumping cash out of Dell until it runs dry.

Dell has already spent $10B+ on acquisitions to transform the company and have admitted those bets have fallen short of their goal, now they’re throwing another $15B into the pot like some down on their luck Vegas gambler thinking that if they up the bet yet again they’re sure to win this time and everything will come good.

That type of guy goes home from Vegas flat broke and with a head full of shattered dreams.

Competitors, beware the ides…