Monday, January 16, 2006

Windows grid: cheaper than you think

As I mentioned in my post yesterday (Grids: All HPC? All *nix? Not anymore), I promised to follow up on one of the arguments used against using a Microsoft OS on a compute grid: cost.

In Kim's eBig talk, one of the audience members mentioned why he thought that Microsoft was behind in grid computing, and he gave a concrete example. He talked about the cost involved in setting up a 256 node compute cluster. If you're paying $400 retail for each copy of Windows Server 2003, that adds $100,000 onto the price of your cluster (even $300 copies of XP Pro would run about $75K). As the audience member observed, that's nothing to sneeze at.

When you compare it to $0 for 256 copies of your favorite Linux flavor, the Microsoft solution does seem expensive.

However, on further reflection, it becomes clear that there are costs beyond OS, and that OS cost alone is not a fair comparison.

First of all, the example assumes a purpose-built cluster. This goes against one of the primary reasons for grid computing: taking advantage of the computers that already exist. If you want 256 Linux boxes in a typical organization, you need to go buy them. But if you need the power of 256 Windows machines--you probably already own them! Because Windows is the dominant operating system, your organization probably has thousands of Windows machines that can contribute to your grid.

Even if you need to buy some new hardware, you can have an "extended" cluster featuring a combination of dedicated hardware (in your cluster) and shared hardware (underutilized servers and/or idle desktops). By using existing hardware, you save not only OS costs but hardware costs as well.

Furthermore, the example only counts the operating system cost--not the cost of any other aspects of the system. Some of the most popular *nix based distributed computing solutions cost one to two thousand dollars per node! Sure, you saved $300 by getting a free OS--then you spent more than triple that on your grid solution. That dwarfs the cost of the OS.

And, it doesn't count what is often the largest cost of all--the cost of setting up the grid. Many of the *nix solutions are what we like to call "thinly disguised consulting projects." They are so complicated that setting them up involves hundreds (or sometimes thousands) of hours of consulting time. Some of the big consulting companies have excellent grid computing practices--but I assure you, they're not cheap. Put a small team of consultants on your payroll at $200 an hour for a couple of months and watch how quickly they, too, outpace your OS costs.

So, are the Microsoft OSs too expensive for compute grids? Probably not. Most likely, you can get an extended cluster using some existing hardware (and OS), some new hardware, add the Digipede Network (at under $200 a node), and set it up without an expensive consulting project.

One more X-factor here: as far as I know, Microsoft has not yet announced pricing for its Compute Cluster Solution, which will be out later this year. Who knows? When that price is announced, it may make the OS decision even more economical...