Analysis What do you get when an $11 billion Windows server business from Microsoft collides with a $10 billion market for supercomputers? You get a chance for Microsoft to get a little extra money for a little extra work, which is about as good as it gets for Redmond these days.
The cross-platform nature of Linux, and its resemblance to Unix - whose many flavors dominated supercomputing a decade ago before Beowulf clustering for Linux on cheap x86 boxes came along - has made it the default operating system for high performance computing in academic, government, and commercial supercomputing labs. And if revenue is any guide, Unix plus Linux still trumps Windows in the data center. But Windows can be - and probably will be - a contender in supercomputing, and this week's announcement of Windows HPC Server 2008 in New York's financial district is but round two in the ring for Microsoft.
Because supercomputing is such a nerdy thing, Microsoft was keen on pointing out all of the features that it has put into Windows HPC Server 2008, which is a cut-down version of the "Longhorn Server" Windows Server 2008 stack that has supercomputing-specific features and tunings added to the operating system.
The feature set goes beyond adding support for the Message Passing Interface (MPI) protocol commonly used for parallel supercomputers and now includes NetDirect, a high-speed networking stack that implements the Remote Direct Memory Access (RDMA) protocol, a cluster job scheduler that is compatible with Open Grid Forum specs, cluster management tools, and fail-over for server nodes in a cluster.
But it may not be such features that give Windows HPC Server 2008 traction in the supercomputing space, even though features are the ante to get into the game. Familiarity with Windows and newbie HPC customers who are willing to pay for an HPC-enabled operating system is what could make Microsoft some money.
For the past several years, the so-called workgroup portion of the HPC space (machines that cost less than $50,000) has been growing considerably faster than the overall HPC server growth rate. In its latest market analysis, IDC said that the global HPC market accounted for just over $10 billion in sales in 2007 and said further that it would grow at a compound annual growth rate of 9.2 percent between now and 2012. The entry segment is growing at about half again as fast as the market at large, mainly because people like to hug their servers and don't like to share them.
As compute capacity gets cheaper, it becomes economically feasible to not use a shared resource. And people get their own machinery, which people like. As HPC applications are ported and tuned for Windows and as HPC users who are unfamiliar with Linux or Unix look to add their own baby supers - particularly those users in life sciences, financial services, and other areas where Unix and Linux are black arts and Windows runs the stuff that mere mortals touch every day - Windows is going to be a popular choice.
Now, here's the funny bit. While Linux dominates the Top 500 supercomputer list, which is a who's who of supercomputing at the largest government, academic, and corporate HPC sites, a lot of the supercomputer centers do not pay for Linux licenses and support. In many cases, these HPC sites know a lot more about Linux and tuning it for HPC than Red Hat, Novell, and other commercial Linux suppliers.
Therefore, they save money by self-supporting. So if you add up all the server nodes out there in the HPC arena, and think about all of the Linux out there, the amount of support money it generates is a lot less than you might expect.
Contrast this with the future entry super buyer looking at taking the easy way out running a Windows-based machine, such as the Cray X1 announced last week. Windows Web Server 2008, the web-only version of Windows costs $469, Windows Server 2008 Standard costs $999 per server with access for five users, and Windows Server 2008 Enterprise costs $3,999 with 25 users.
The $475 per server cost for Windows HPC Server 2008 seems like a fair price by comparison, particularly when Windows Vista Business or Ultimate editions - which cost $300 and $320 respectively - don't know jack about HPC clustering. A Windows workstation with two processor sockets and maybe eight cores - that takes forever to do a job - is the choice these entry HPC customers are looking at. That extra cost for software seems minimal if it enables a baby cluster that can do a lot more work for the money.
And, this is money that will not go to a Linux vendor, because Linux is not even an option for this customer. They want to do algorithms for folding proteins, for analyzing gene sequences, or for Monte Carlo simulations to make stock picks.
Even if such customers get Windows HPC Server 2008 for half price from Microsoft and its resellers, this is more money than they would have paid for Linux. In a way, the fact that Microsoft is able to charge $475 is a measure of the barrier of adoption for Linux at the emerging low-end of the HPC server racket. People will pay not to have to learn a new thing.
Now, having said that, it will take forever and a day for Microsoft to make a lot of money in HPC. At $475 a pop, Microsoft has to sell the operating system on 2.1 million server nodes to make $1 billion. The world only consumes a little more than 8 million servers a year. So it is not like the HPC variant of the Windows stack is going rake in billions and radically expand the Windows share of the worldwide server pie. But every dollar spent on Windows in HPC is one that a Linux vendor did not - and maybe never could - capture. And that is worth something to Microsoft. ®