Companies continue to buy a lot of server computers to run their applications and web sites and perform other routine computing chores. More than 7 million servers were purchased last year, according to estimates released last week by analysts, generating about $50 billion in revenues for server manufacturers like HP, IBM, Dell, Sun, and Fujitsu. The server market is a big one, and it looks fairly healthy at the moment. But that may be an illusion. There are growing indications that the server business is a dead man walking.
The most immediate threat is the twin trends of consolidation and virtualization. To save money, companies are merging their data centers and standardizing the applications they run throughout their businesses. The chemicals giant Bayer, for instance, has been consolidating its IT assets worldwide. In the U.S. alone, it slashed its number of data centers from 42 to 2 in 2002, in the process cutting its server count by more than half, from 1,335 to 615. With more companies embracing server virtualization – the use of software to turn one physical server into many virtual servers – the opportunities for further consolidation will only expand. Last year, for example, Sumitomo Mitsui Bank used virtualization to replace 149 traditional servers with 14 blade servers running VMware virtualization software. Timothy Morgan, editor of IT Jungle, believes that, as companies accelerate the consolidation and virtualization of their computing infrastructures, “the installed base of 20 million to 25 million servers in the world could condense radically, perhaps to as low as 10 million to 15 million machines.”
But even that may, in the long run, be too rosy a scenario.
What if the future brings not simply a rationalized version of traditional computing, with fewer servers used more efficiently, but a fundamentally different version of computing, with little need for brand-name servers whatsoever? In this scenario, the core unit of business computing would not be small, inflexible servers but rather large, flexible computing clusters or grids. These clusters in turn would be built not from traditional branded servers but from cheap, commodity subcomponents – chips, boards, drives, power supplies, and so on – that the grid operators would assemble into tightly networked physical or virtual machines. Many of the functions and features built into today’s branded servers would be taken over by the software running the cluster.
If you want to see a harbinger of this model of computing, just look at Google’s infrastructure. Google doesn’t buy any servers to run its search engine. It buys cheap, commodity components and assembles them itself into vast clusters of computers that it describes as “resembl[ing] mid-range desktop PCs.” The computers run in parallel, using a customized version of the open-source Linux operating system. Google doesn’t have to worry about “server reliability” – one of the main selling points used by server manufacturers – because reliability is ensured by its software, not its hardware. If, say, a processor fails, others pick up the slack until the faulty part is swapped out. What concerns Google is the big cluster and the little subcomponent; it’s moved well beyond the idea of the branded server being the heart of business computing.
Obviously, Google is a unique company with idiosyncratic computing requirements. But its efficient, flexible, networked model of computing looks more and more like the model of the future. As Google engineers Luiz Andre Barroso, Jeffrey Dean and Urs Hoelzle write in their IEEE Micro article Web Search for a Planet, “many applications share the essential traits that allow for a PC-based cluster architecture.” IT expert Paul Strassmann goes further in arguing that Google’s infrastructure serves as a template for the future. “Network-centric systems,” he says, “cannot be built on [traditional] workgroup-centric architecture.” If large, expert-run utility grids supplant subscale corporate data centers as the engines of computing, the need to buy branded servers would evaporate. The highly sophisticated engineers who build and operate the grids would, like Google’s engineers, simply buy cheap subcomponents and use sophisticated software to tie them all together into large-scale computing powerplants. Or, if they wanted to continue to purchase self-contained computer boxes, they’d simply contract with Taiwanese or Chinese suppliers to assemble them to their specifications, cutting out the middlemen and their markups.
Ultimately, we may come to find that the branded server was simply a transitional technology, a stop-gap machine required as the network, or utility, model of computing matured. I recently spoke to the chief executive of a big utility hosting company who expressed amazement that its largest server supplier seemed to be “in denial” about the profound shifts under way in business computing. Maybe it is denial. Or maybe it’s just fear.
UPDATE: See further discussion here.