I am in an “Economics of Strategy” class as part of a part-time MBA. We expect to put forth a thesis that Microsoft should invest in a vertical integration into green data centers as a means of preserving then replacing software revenues current at threat under the cloud.

Imagine our glee with the Wednesday announcement of the large-capacity MS/Chicago Data Center. Probably not so green yet until they get to rural and renewables.

Can I assume your permission to reference your methodology as a means of explaining the market to our non-technical audience?

Certainly, scale and management are necessary but not sufficient for cloud infrastructure. For any infrastructure reliability, availability and serviceability are also vital, among many other attributes. But when you look at cloud infrastructures – or those that would be cloud – high scale and a degree of autonomic management seem to be key enablers. To your point about logistics ecosystems, this is all plumbing, and whether we go to a utility model or a retail model or something else for delivery is still up in the air.

Robin

I believe this is the defining aspect of Cloud Computing. Out of necessity the first players have built their own vertical clouds, but the future of Cloud Computing is not a forest of chimneys. Clients will use a variety of services, some specialized for their line of business, some fairly general, from a dozen or more suppliers. Just as our physical logistics ecosystems allows us to purchase goods, from specials to commodities, from a set of stores in our neighborhood, the cloud ecosystem will have to support the frictionless exchange of digital services and their data to support the whole breath of the market. Scalability and management are necessary, but not sufficient elements to allow a storage system to flourish in the cloud.
Ernst

Just read this post now, thanks for the shout out I’d humbly like to offer up (our partners) Newservers.com as a provider, but I am not quite sure where to put them.

We have a nice ~0.2PB HPC storage cluster going into a supercomputer site soon. Benchmarks I’ve run yesterday on one node are at ~3GB/s sustained (no SSDs in this apart from the boot drives … all spinning rust) for streaming 192GB from disk, 2.5GB/s streaming it to the disk. Running GlusterFS atop Infiniband QDR atop our boxen. Pretty seamless incremental scalability … drop another node in, boot it up, and about 10 minutes later its available. GlusterFS doesn’t (yet) have the ability to hotadd/remove storage, but it is in the roadmap for next release. Until then, we are automatically generating the requisite server config files, and can automate a push or a cause a pull and restart on the client side to provide an approximation to this capability.

My point being that GlusterFS should definitely be up there, certainly in HPC, but they are doing some interesting things into the cloud space. Might be worth a talk with Hitesh Chelani about their offerings.

As for Scalable, we are small, but I can say that interest in our offerings is growing fast and hard. The last several months of ours have been some of our best in the company history. We now have value added resellers integrating/using these systems in their own offerings, as well as customers/partners building larger (cloud-like) storage.

Our placement on the Y-axis, however … does not given enough credit to the scalability of Isilon and OneFS. You define scale as “single namespace; single management domain; reasonable upward performance increment per incremental dollar investment.”

Single namespace and single management domain – from as little as 6 TBs to over 5 PBs. Check.

Single system performance which scales up to 45 Gbps and a 144 nodes – with the customer being able to choose, on a per-node basis, whether they are staying along the linearity curve by adding like nodes or deviating from it. Our customers can add denser nodes over time, which adjusts the price/performance curve towards lower cost or they can add higher performing nodes (and accelerator nodes) and increase the performance over time. I would certainly argue that the cost per node is fixed/predictable and competitive with any other commercial vendor above us on the scale axis. Check.

Not only can customers dynamically adjust their system over time to maximize cost or performance, they can do so in a fully backwards-compatible, investment maximizing way. An Isilon customer can happily run multiple generations and types of hardware in the same cluster and choose how and when to adjust – without any disruption, of course.

All of this with a highly advanced protection model incorporating Reed-Solmon encodings – allowing for up to four simultaneous failures with 20% overhead or less and /per-directory flexibility/; core enterprise features – including per-directory snapshots, quotas, replication; and of course, a proven track record – over 1000 customers, over 100 PBs installed, and some of the world’s leading social networks, next-generation cloud bioinformatics, photo-sharing sites, etc.

There is no doubt that Isilon, based on a proven track record of highly scalable technology and demanding customer environments, should be firmly in the upper-right quadrant.

Nick
http://twitter.com/isilon_nick

Robin, I’d be interested in your scoring methodology for this chart. For example, why did you score Parascale as more cloud like than ByCast? Why is Atmos worse than all the others? Why is MaxiScale so far into the upper right corner? My interest is not academic: we’re currently intensely interested in cloud storage technologies, with a petascale-level problem lurking immanently nearby…

Joe Kraska
BAE Systems
San Diego CA
USA

The other thing to consider in this argument, as it pertains to PAYG and multi-tenancy, is how X-axis “management” is set up to actually support both those objectives.

I’d agree with the statement that they are binary elements which you either have or don’t have. However, some of those cloud storage technologies are specifically set up to support keeping track of and metering thousands of unique customers. Others are less service-provider friendly.

It’s great to have an “autonomic” system that saves you opex on storage admin salary – but if you’re spending all that opex on software programmers, billing analysts, and account managers, the opex goes out the door all the same.

I think what muddies the water here, as happens with many cloud discussions these days, is the merging of “Internal” cloud with “Service Provider” cloud. It’s understandable why they merge – after all who wants to exclude one customer base in favor of the other? – but evaluation against the X-axis is different depending on your scenario.

Another vendor worth evaluating in this space is DDN – their Web Object Scaler (WOS) technology is upper-right material.

I have trouble with the Cloud nomenclature: I simply don’t think of Isilon, MatrixStore, etc as Cloud storage, nor do I think of Cloud storage as being “location transparency”.

For me, the term cloud storage strongly implies that there is a managed (or self-managed) soft service available to clients. Sure, the underlying components that make up that cloud based service may contain a set of the hardware solutions listed, but the premise that the cloud quadrant contains storage solutions as opposed to storage services (of which I would list: storage services (aka s3), HPC, applications (aka google docs), etc) doesn’t sit well with me… I’d rather call the chart a “Scalable Storage Solutions” chart, or the such like. Obviously that would defeat the purpose of “bringing clarity to cloud terminology”, but is this really a cloud quadrant?

However, that being said, I find the axes thought provoking and a very useful/visual way in which to compare competing products in a marketplace that has so many shades of gray.

Sun QFS and SAMFS open source like ZFS. Interesting road map on all of them.

Other interesting open-source here
MogileFS, Hadoop, Coda, Sun Celeste

Wrong understanding! Even our own in-house testing is on over double that amount of nodes, with customers going well beyond that. We have no way of verifying the exact number of nodes as we sell by the Terabyte – of which some sites have over 500 now.
However, that is nothing compared to a few projects under preparation. One of them calls for 1,500 servers – 15,000 disks in total – 15 Petabyte – all of that in a single flat namespace (we don’t do FS as you may recall). We’ll keep you posted.

–Paul

BTW, it would be nice to apply equal rigor to all names on the chart – including some of the stuff in the upper right corner. I understand clouds consist mostly of vapor, but still…

My understanding is that Caringo hasn’t tested and/or installed over 32 nodes in a single FS. Update: Paul tells me that they’ve tested over twice as many and that customers have gone well beyond that – but because they license by the TB and not the node they don’t have good visibility into how many nodes their bigger sites have. End update.

Robin