Computer system scalability: Up or out?

Business starts with a particular set of computing needs. If it grows, the organization needs to be concerned about the scalability of its computer systems.

Scalability is a good thing, because it frees you from having to start over and build a completely new system with new software and hardware. With a scalable system, you will likely keep running the same software and merely add more hardware. There’s nothing new about this idea, but the arrival of e-commerce has changed the way businesses think about scaling their computer systems.

Two terms characterize the strategies for scaling: scaling up and scaling out.

Scaling up is the traditional approach: Instead of a small server, get a bigger, multiprocessor server (or a cluster of them). If that’s not enough, maybe you need either a mainframe or a supercomputer.

One problem with this approach is that it takes time. In today’s e-commerce environment, long-range planning and efficiency may matter less than being able to bring a new application online quickly. Therefore, you might want to try scaling out, which is a relatively fast and easy approach that involves buying a bunch of inexpensive, commoditized server boxes.

Neither of these approaches is inherently better than the other, with each having its own advantages. And today, e-commerce firms are adopting pragmatic growth strategies that call for both scaling up and scaling out. As the number of transactions and site visits increases, Web servers and application servers are being scaled out with the addition of other small servers that run multiple copies of Web or application software. Meanwhile, back-end database servers are being scaled up in a slower, more deliberate way, with large, multiprocessor servers running a single copy of the database software.

Brian Richardson, an analyst with Stamford Connecticut-based Meta Group Inc., says it’s feasible to scale out Web and application servers because there’s no need for each transaction to take place in a single copy of the application. Database servers, however, work better when running a single instance of the application. This prevents transactions from running afoul of one another – for example, a retailer doesn’t want to promise delivery of an item that isn’t available in inventory.

Richardson says the “rack-and-stack approach” of scaling out Web and application servers with one- or two-processor units is both simple and cheap. “The big driver here is not total cost of ownership and efficiency, from an operational perspective,” he says. “The more important driver has become time to implementation – how quickly can you get the new application up and how quickly can you do business-to-business e-commerce with your constituents?

“Total cost of ownership takes a back seat to flexibility and adaptability,” he continues. “If you can implement small servers easily and quickly and save six months in time to implementation, it’s worth it.”


Wayne Kernochan, an analyst at Aberdeen Group Inc. in Boston, agrees.

“It’s fair to say that using multiple small servers is a quick and dirty approach,” he says. “With a new Web site application like e-mail or directories, you just slap in another server. … When you’re in a tearing hurry to get started and the long-term future doesn’t matter, this approach makes a lot of sense.”

However, David Friedlander, an analyst at Giga Information Group Inc. in Norwell, Massachusetts, says that by scaling up instead of out, “you get a large quantity of processing power devoted to a single instance of an application. If you are running a stock exchange or processing thousands of financial transactions per second, then you need that more powerful processing engine.”

“On a single large server, you don’t need to consider the network as much because it’s inside the box, and that makes administration simpler,” adds Kernochan. “That’s important, because our studies show that administrative costs tend to crowd out other costs in the long term.”

But in the short term, it’s the cost of purchasing, rather than maintaining, the server that matters most to e-commerce companies. Server prices vary, but Kernochan says an eight-processor machine can cost twice what you’d pay for a two-processor unit. The pricing equation doesn’t shift in favor of the larger machine until you have scaled out beyond four processors or when your database grows too large, he says.

In light of the continuing rapid growth of e-commerce, it’s likely that each strategy will continue to have its place.

Alexander is a freelance writer in Edina, Minnesota.