New NC problem: traffic

The challenge is to come up with server engines that can provide good performance for the hundreds--if not thousands, in some cases--of NCs being deployed within a single company.

Unlike a personal computer, the NC relies on a server as the computing engine to do much of the information processing. But this can be a daunting task and can bring typical servers to their knees when hundreds of NCs are trying to work with servers simultaneously.

To fill this need, vendors like IBM (IBM) and Tricord Systems unveiled strategies this week geared to provide the engines that will drive NCs. (See related story)

IBM this week positioned its powerful RS/6000 SP server platform as a potential solution. The SP server can bring to bear a "cluster" of dozens or hundreds of server "nodes" to handle all this NC traffic. In advanced clustering technologies like IBM's, these server nodes are strung together to form a giant virtual computer.

"In the NC space, clustering is better than [traditional multiprocessing]," says Tim Dougherty, manager of network computing solutions at IBM's RS/6000 division.

Dougherty claims that there is too much overhead involved in making all the processors work together in large-scale "symmetric" multiprocessing, one of the most popular forms of multiprocessing. This overhead typically leads to multiprocessing systems that do not scale up in performance efficiently as more processors are added, he said.

Typically, this kind of system uses one copy of the OS and runs it across many processors. With clustering, on the other hand, each server has its own OS copy and talks to the other servers in the cluster through a special communication channel.

"You'll wind up paying more [for IBM's RS/6000 node clustering], but it has a management advantage," Dougherty added.

Each server node consists basically of an IBM POWER2 microprocessor or PowerPC processor, memory, and a large-capacity hard disk drive storage system. A system of up to 512 nodes can be constructed, and each node is managed by a copy of the AIX operating system.

The system can also be partitioned into "pools" of nodes, according to IBM. For example, two nodes can work as a Lotus Notes server, while ten others process a parallel database.

Depending on the application, performance may not be the only reason to consider clustering, according to Jim Fulton, director of product marketing for Network Computing Devices (NCDI). NCD manufactures IBM's Network Station.

"I think that the first consideration is the structure of the organization. In many companies, there may be a need for specialized applications at the department level" that are not used in other departments, he says. In such cases, clustering is a logical solution to make management of the system easier as well as increase performance.

For smaller deployments, use of SMP servers may still be appropriate. Fulton says their WinCenter software, which is designed to efficiently send Windows applications, graphics, and audio from the server to the client over an intranet, can allow 20 or 25 people per processor for Windows NT servers that use Intel Pentium processors.

"Those things [Windows terminal] have been around a while, but now the concept is getting everyone's attention, and it has worked quite well with small departments and 20 to 25 users, even when connected remotely over modem," said Dave Cappuccio, vice president of research for the Gartner Group research firm.

Still, one thing is becoming apparent. While the NCs are being touted as low-cost alternatives to the PC, the hidden cost of servers is just coming to light. This surely will have considerable impact on many an IS manager's plans to adopt the NC.