User demands force rethinking of net CPUs

User demands force rethinking of net CPUs
Fundamental changes are coming in the way the Internet is used and how consumers access and navigate the Web. The upshot is that designers of server hardware and software are being forced to rethink their architectural assumptions, all the way down to the chip level.

Changes with the most profound impact on server hardware and software design fall into four broad categories: the still-increasing numbers of users and types of devices accessing the Web; the higher-bandwidth connections available; the new object-based methodologies being developed for Web services; and the merging of traditional server clusters and telecommunications/data communications into a single Internet Protocol-based topology.

To date, in a still PC-dominated Web user environment, the number of users who access the Web for an average of 20 hours a month for up to two hours at a time has increased from 24 million in 1998 to about 180 million in 2001, according to Nielsen/Net Ratings Inc. Adding to the pressure on server designers is the rush to high-bandwidth access by many of these users. According to recent surveys by Forrester Research and Nielsen, almost 20 percent of all households, about 13.8 million, have high-bandwidth access. It seems clear that the usage patterns can be expected to skyrocket regardless of market conditions.

The server community has long been recognized, according to Greg Brashier, vice president of business development at Strategic Research Corp. (Santa Barbara, Calif.), and this is the reason industrywide and company-driven efforts such as Infiniband have gained traction in the market.

But with the higher bandwidth available within and among clusters of servers, the ways in which service providers want to build their computer systems has changed, said Kevin Dierling, vice president of marketing at Mellanox Technologies Inc. (Santa Clara, Calif.).

The original hierarchical environment of clusters of several dozen traditional servers supported by dedicated appliances and Internet service providers is being simplified for maximum flexibility and scalability using configurations of hundreds of server "blades." These are small, board-level full-function servers clustered together to provide particular Web site services, including storage-area networking access, I/O processing, e-mail, specialized media handling, security and firewall operations.

"Since I/O bandwidth is less of a problem, it is no longer necessary to congregate various Web services on a single server or cluster of large servers," said Dierling. "Now enterprise data centers and Internet service providers can have much more fine-grained control."

In addition, the nature of the job required of these nodes has changed drastically over the past few years, said Marc Erickson, project manager at IBM/Object Technologies International Inc. (Raleigh, N.C.). This is due to the proliferation of Web-capable information appliances.

One of the first areas to be modified by this changing usage pattern is server software and operating systems. According to Erickson, specialized application server software has had to be developed to handle the interactions with the literally millions of small-footprint "pervasive-computing" devices that are proliferating. In the new Web services environment the nature of the workload is changing from strict HTML page accesses to much more complex interchanges.

Designers of the basic processors used in servers are rethinking their designs in several fundamental ways, mostly in a shift to sophisticated multithreaded architectures that more closely match the workloads servers are, or soon will be, handling.

See related chart