Headhunter.net is also trying to get its hardware closer to its customers, setting up servers in Georgia, Virginia, and California. This created another issue--how to balance loads across geographic distances. Headhunter.net chose WSD-NP, a $14,000 load-balancer from Radware Inc. that evens out loads between Virginia and California servers and sends any overflow to Georgia. Other products can balance loads between processors in a given rack, but this was one of the few devices that balances loads between geographically dispersed servers, says Mark Fouraker, VP of operations.

Load balancing, a feature of clustering, is a common strategy for boosting server performance. A large task is broken up and spread among servers. This speeds processes and may even create functionality that's not otherwise possible. "Clustering allows you to solve a problem that isn't solvable with just one server," says David Gelardi, director of benchmarking and application performance for RS/6000 at IBM. As an example, Gelardi says a cable company must have its billing completed during a specified time cycle. If a single server can't do the job quickly, the company doesn't have the luxury of waiting. To boost performance, the company will use clustering.

Susan Blount, manager of enterprise server product marketing at Dell Computer, is an advocate of adding cache. According to her, the cost of a processor can increase from $750 to $1,500 when doubling the cache from 500 Kbytes to 1 Mbyte, but the added cache will increase server performance by up to 20%. "On the desktop, cache doesn't do much," she says, "but in servers, people underestimate the importance of cache and instead focus on processor speed."

CPUs Count
That's not to say processors are unimportant. CPUs can boost server speeds in two ways: by adding more processors and by swapping for faster ones. Pertti Manner, group marketing manager for Sun Microsystems, cautions that users who want to boost performance by adding more or faster CPUs must be sure their system architecture can handle this new hardware.

Customers make decisions about CPUs based on the processor's age, Manner says. "A customer who has had a system a short time, like six months, and who needs more performance, will often buy more CPUs. But if that same customer bought a system two years ago, they exchange their existing CPUs for faster ones." The rationale for this is that in the short term, CPUs don't change much, so more CPUs are needed to boost performance. Over the longer term, new CPUs can provide a significant performance boost while not requiring more processors.

About 10 months ago, Tenent Healthsystem boosted performance by changing its operating system so it could add CPUs. It moved from an early version of Novell's NetWare, which supported one CPU per server, to NetWare 4.1, which supports two. "We increased users from 450 to almost 750 in one year," says Jeff Lett, director of telecommunications and technology. He's looking to boost performance even more by setting up a centralized storage capability. "We're looking toward a storage area network where I'm not relying on individual server performance but on cluster technology," he says.

Garrett Grainger, VP of information services at Dixon Ticonderoga Co., a manufacturer of writing tools and office supplies in Heathrow, Fla., added servers when customers began to use its Web site in increasing numbers to access information such as order status. He now has about a dozen servers and uses clustering technology based on proprietary software.

Grainger likes adding RAM, with each of his servers now having at least 1 Gbyte. "The bigger the memory, the better the response," he says. Grainger is in good company. A recent survey of NT system managers by American Business Research found that two-thirds believed RAM was the most significant factor in server and workstation performance.

continued...page 3
return to page 1

Photo of Fouraker by Reid Horn
Photo of Lett by Mark Escher