The coming of the megacomputer
March 06, 2009
Here's an incredible, and telling, data point. In a talk yesterday, reports the Financial Times' Richard Waters, the head of Microsoft Research, Rick Rashid, said that about 20 percent of all the server computers being sold in the world "are now being bought by a small handful of internet companies," including Microsoft, Google, Yahoo and Amazon.
Recently, total worldwide server sales have been running at around 8 million units a year. That means that the cloud giants are gobbling up more than a million and a half servers annually. (What's not clear is how Google fits into these numbers, since last I heard it was assembling its own servers rather than buying finished units.)
Waters says this about Rashid's figure: "That is an amazing statistic, and certainly not one I’d heard before. And this is before cloud computing has really caught on in a big way." What we're seeing is the first stage of a rapid centralization of data-processing power - on a scale unimaginable before. At the same time, of course, the computing power at the edges, ie, in the devices that we all use, is also growing rapidly. An iPhone would have qualified as a supercomputer a few decades ago. But because the user devices draw much of their functionality (and data) from the Net, it's the centralization trend that's the key one in reshaping computing today.
Rashid also pointed out, according to Waters, that "every time there’s a transition to a new computer architecture, there’s a tendency simply to assume that existing applications will be carried over (ie, word processors in the cloud). But the new architecture actually makes possible many new applications that had never been thought of, and these are the ones that go on to define the next stage of computing." The consolidation of server sales into the hands of just a few companies also portends a radical reshaping of the server industry, something already apparent in the vigorous attempts by hardware vendors to position themselves as suppliers to the cloud.
"every time there’s a transition to a new computer architecture, there’s a tendency simply to assume that existing applications will be carried over"
Case in point - the current-form-spreadsheet. With the massive amount of data available to business users in the cloud architecture - Datamining techniques like Clustering , Decision Trees , Network Analysis will be available as features and the current charting capabilities would really move to the world of visualization.
more here http://www.gandalf-lab.com/blog/2009/03/how-many-computers-does-world-need.html
Posted by: niraj j at March 6, 2009 03:44 PM
I will cite our old Luddite friend Marshall McLuhan, who said that each new media fails to achieve the goals it was created for, but recapitulates the media it replaces and brings those old goals to successful completion. In this case, clouds and grid computing are still a solution in search of a problem, except in one area: render farms. Some of the largest server computer installations in the world are render farms at graphics studios like Pixar, ILM, WETA, etc. These racks of servers run applications that were previously done by supercomputers, workflows that were originally designed for supercomputers have been completely reworked to adapt to the server/client grid computing model.
I remember visiting Digital Productions in the 1980s when they were struggling to make enough money to pay the electric bills for their Cray-XMP. Today, companies like RenderCore operate render farms that can be rented by the minute. These on-demand supercomputer/grid rentals can succeed because they can serve multiple customers and keep a queue of work flowing, while previous supercomputer systems needed single big-money clients and tended to die off once the big job was done (like DP did).
Posted by: Charles at March 6, 2009 04:24 PM
Yeah, but, watch those numbers in 2009 and 2010 as two factors set in: the economy and better technology.
Everything I hear leads me to believe that Google et al. are having a lot of trouble monetizing what they've already got nevermind what they're building that's in the pipeline. Rather, I think this is a poker game gone wrong:
Investors (and analysts) bought the hype these firms came up with: Yes, commodity computing is coming. The winning position is to be there with capacity to meet demand and with initial market share and momentum to obtain lock-in on the platform. Therefore, let's build-build-build. Data centers, containerized computing, all that.
It was a poker game in that, not only are these firms gambling on their ability to define the platform and pick the right technology, additionally they were trying to bluff each other and the markets about the best build-out rate. So, for example, company G starts building data centers, then company M starts. G notes that M has made hardware choices that G is just about certain are wrong. G knows that nobody at the table *really* knows when this market takes off. G bluffs by telling everyone that the market takes off any day now and starts talking very loudly about all their new build-out -- in hopes of getting M to double and quadruple down on what G sees as M's losing bet. Of course... M is doing exactly the same thing to G: the pot is going to go to the high card (a spot card, no less) cause nobody's got anything in their hand.
Meanwhile, all of this looks plausible enough to the analysts that, for a time, investors reward this behavior.
Google's got some hedging in there - clever hedging. They're hedged in the real estate, in the sweet tax breaks various sites were built with, in the (very flexible, apparently well-thought out) physical design of their facilities, in the prison-like security they've built around these death-camps-in-waiting, and in the power and water rights and facilities they've secured. There's a fairly high lower-bound on their returns here (but still, likely, a sub-0 lower-bound on their profits).
The computing systems, though: they're all wrong. Here is where the thing unwinds and makes a mess for these firms:
The build-out curve they are implementing starts with a very steep upwards trend, probably starting to level out now as they respond to the lack of demand and generally failing economy. So, that means that by today they own a heck of a lot of theoretical capacity - actual machines. More than they have easy use for, on a reliable basis.
Meanwhile, the hardware technology curve continues to plod-along as a "moore" kind exponent (really, as a kind of fractal that is everywhere locally an exponential curve). So, the firms have all those physical machines, under-used, and the hw tech is shifting out from under them.
Meanwhile, the really interesting one, the software technology curve is going through an accelerated phase in a pattern of punctuated equilibrium. Sometimes software just incrementally plods along in improvements, for a while. Other times, in a very few short years, suddenly entire classes of new kinds of software are introduced, often turning older stuff into relics.
Commodity computing was academically not that big a deal until the past few years. Now, it's an m-f-ing huge deal. The prior neglect means that there's lots of low-hanging fruit. The current high level of attention therefore indicates that the software technology curve in commodity computing is likely to be quite steep for, say, the next 10 years.
Therefore: like hw, software is shifting out from under the sunk costs of the big-firm data centers.
Software is "really interesting" here because the catalyst for an eventual commodity computing demand spike will be equal parts "immediately useful applications" and "a platform that is compelling in the long-view". Software is what creates that platform. The applications are new software. And in these areas: software is changing very, very rapidly.
In a few years we'll have a much more solid idea of what the commodity computing platform will look like and, odds are, it will look very, very different from what runs on servers today. The software itself will likely be designed from the ground up to save on hardware complexity and power consumption (e.g.,: there is no obvious need for hardware-supported virtual memory).
Initially, such new platforms will of course run on today's server hardware but the weird thing is that it will be very cheap and easy to make simpler hardware that runs the new platform far less expensively.
At that point - and remember - this is just a few years out by current trends - today's *speculative* poker-game data centers are going to be white elephants. Very pricey white elephants whose associated liabilities make it hard for these big firms to compete with the new hw and new platform.
I only very vaguely know a bit about the history of rail-road build-out and expansion in the US, or telegraphy, but, didn't they go through some similar phase? E.g., over-exuberant speculators climbing over one another to lay down tracks of an ultimately losing gauge or paying too much for desirable but not-as-critical-as-hoped right-of-ways.... that kind of thing? Didn't some big money lose their shirts?
I suspect a similar fate awaits today's commodity computing big movers with some exceptions. The most well-known exception would be Amazon. I don't see Bezos as the next Vanderbilt but AMZN seems unique in taking a more or less "pay as you go" approach to build-out so they can't exactly lose very hard.
I wonder who will be the Astors, though? Whoever it is, I doubt they've bothered making any serious moves among the current efforts. They'll swoop in with financing after Google, MSFT, the taxpayers, etc. fund the R&D and the new-Astor's will win since they won't have the sunk costs and liabilities of those firms.
In the alternative to all of the above: well, in that case I'm just completely wrong. :-)
Posted by: Tom Lord at March 6, 2009 05:47 PM
The shift in server sales toward the cloud-builders could be clearly seen last spring, when HP, Dell and IBM all suddenly announced "cloud server" products that focused on density, power management and half-depth servers. This was an area that had previously been the focus of folks like Rackable. It's not an accident that those companies also introduced container products on the heels of Microsoft's announcement that it was shifting to containers.
There are no more "speculative, poker-game" data centers. Microsoft and Google have adjusted their timetables to build out the announced projects only when existing data centers approach capacity. In the current environment, any data center that doesn't already have tenants is simply not being built.
Posted by: RichM at March 6, 2009 06:55 PM
Yes, that's what I referred to as a "flattening" of the initially steep build-out. There was a poker-game for a few years but reality has intruded - as I said. They are left with white elephants. Nick's piece re-reports and analyzes the impact of this transpiring on server sales, which is an interesting angle. If you look at what these firms were saying a couple of years ago, this retreat in the vague direction of "pay as you go" was not in their cards. I'm not sure they really have that in the draw here, either. Like I said: pot will go (among these firms, Amazon excepted) to the highest card - a spot card. They (especially Google) have more than they're going to have use for already.
On a more serious note and not just to RichM:
The financiers have really screwed the pooch in this industry because what they call "horizons" are really "engineering constraints" and they've basically suppressed a lot of useful work for about 20 years now and built out to spectacular proportions a lot of unsustainable work. Lesson for them: controlling money doesn't make you an engineer - it only means you really ought to become one lest you cause great and widespread human tragedy.
It's a lesson arriving too late, of course.
Posted by: Tom Lord at March 6, 2009 07:19 PM
An interesting article and some very interesting comments.
This reminds me of the battle of the ISPs. The encumbent ISP had to build a bunch of special equipment, a bunch of custom hardware and be his own systems integrator. Next the upcoming new ISP wannabe comes along and buys half the number of servers the encumbent bought at half the price the encumbent paid (better off by a (conservative) factor of 4.
But wait, it gets better. While the encumbent had to buy customer ports in groups of 24 (or whatever - pick a number) the newbie ISP buys a single box that can handle 500 ports. OR simply buys ports from some newly establish "port" provider (whether phone or DSL - it does not matter for the purposes of this discussion.
So - bottom line - the newbie to the "plate" has a considerable advantage, in terms of capital equipment outlay and ease of ownership/reduced cost of operation. By at *least* a factor of 4. And, likewise, his successor newbie, will enjoy a similar advantage - making it very difficult and costly for the original (encumbent) ISP to remain profitable.
Likewise you'll see the same "arms race" with cloud computing - the newbies will gain a considerable advantage - able to replace hundreds of older generation machines with newer, cheaper, more performant boxes. [look at the upcoming gen of dual Nahelem machines versus older Xeon based boxes!]
But .. the real requirement for cloud computing is IMO, not computing "horsepower" in terms of CPU speed/# of cores etc. - but simply: addressable memory space. IOW, I'm saying that if you buy *yesterdays* servers and max it out with the affordable memory parts you *might* have a 32Gb machine. But, tomorrows machines will have 256GB (or more) in a single box - and *that* is what will give the newbie the ultimate advantage. And that is also what the established, encumbent, provider will not be able to match - without going to a lot of trouble and writing a bunch of specialized software that is expensive to develope and expensive to maintain. OR - if you look at it another way, assuming that the old/newbie players spend the same dollars on their software, the newbie will end up with 8 times the memory space for the same software $ cost and that will allow them to deliver that much more performance - however you wish to measure "performance". You could say that it allows them to support 8 times the number of simultaneously active users or you could say they can provide instant access to 8 times the amount of live data - or you could say it allows them to support 8 times the I/O measured in terms that the users understand - as useful I/Os per Second. Of course the answer will really be a combination of all of the above - but my point is obvious - the newbie will enjoy a considerable advantage at a much lower cost of ownership.
in these exiting times of transformation of network architecture one can already observe a "dialectical" development: watch the rapid growth of harddrive sales in the end consumer markets used for external local storage like in homeservers and nas devices - or the rise in sales of cheap virtual servers for end consumers used for p2p, gaming, blogging and messaging. the p2p, or end-to-end paradigm will not be simply replaced by the cloud paradigm but give place to surprising new models of computing and distribution.
Circa 2001 I helped maintain what was arguably the most comprehensive map of datacenters, tenants and connectivity available worldwide.
What you are saying has been foretold for many years, it's not that incredible.
Building big datacenter next to hydroelectric dams; Goog building their own servers and common understanding that of course a handful of companies are buying big iron aside, cloud computing isn't any good unless the apps are useful and secure.
Google has trouble keeping Docs secure and this week one of Media Temple's hosting grids went down due to a vendor's hardware failure.
Besides Salesforce, where is the consumer-facing success in the cloud at this time?
Posted by: relaxedguy at March 7, 2009 10:00 AM
Post a comment
Thanks for signing in, . Now you can comment. (sign out)(If you haven't left a comment here before, you may need to be approved by the site owner before your comment will appear. Until then, it won't appear on the entry. Thanks for waiting.)
"Riveting" -San Francisco Chronicle
"Rewarding" -Financial Times
"Ominously prescient" -Kirkus Reviews
"Riveting stuff" -New York Post