To prove that Intel-based technology can drive major back-office applications, PC server vendors are designing their own symmetric multiprocessing (SMP) and clustering technologies to scale Wintel ever higher.
Meanwhile, differing strategies from IBM and Dell Computer highlight the brewing religious war over which server technology scales best.
IBM will offer a clustering solution next quarter and a 16-way and higher SMP solution in the Merced time frame (midyear 2000.)In the meantime, Dell will move beyond two-node clustering for fail-over this summer, but plans to stop short of offering anything beyond eight-way SMP solutions.
Code-named Cornhusker, IBM's technology will move clustering beyond the current Windows NT limit for two-node fail-over. Cornhusker will use IBM's Phoenix technology, which is used to divide tasks among multiple servers within an application.
Cornhusker, built on top of Microsoft Cluster Server, or MSC, clusters can accomplish an eight-node cluster for load balancing, according to an industry source familiar with IBM's plans.
If one of the nodes in the cluster fails, any of the other seven nodes will pick up the balance, according to the source.
Intel-based server vendors realize that IT organizations want rock-solid reliability before deploying mission-critical applications.
"When you start driving applications like SAP, customer tolerance just plummets," said Jonathan Eunice, an analyst and IT adviser at Illuminata. "Everybody is struggling to get big system attributes onto little servers. The price performance of an Intel system is very compelling, but customers don't want classic Intel server availability."
Given a choice between sheer horsepower or availability, vendors now realize that many IT organizations would choose the latter every time.
"Availability is king today. IT managers would be more than willing to sacrifice 10 percent in performance if they could be assured greater availability, and clustering is the only way I know of to get true 24-7, 365 availability," said Richard Sims, senior marketing manager for the server group at NEC.
Following the introduction of Intel's 64-bit Merced chip in 2000, IBM will shift its focus onto SMP boxes. Big Blue will ship its own chip set, code-named Summit, in Netfinity servers along with Intel's 64-bit McKinley processor, due in 2002.
Summit will allow a system to address as much as 1 terabyte of memory and scale to a minimum of 16-way systems, according to Randy Groves, vice president of Netfinity development at IBM.
Dell will be making a major move to scale its PowerEdge server capabilities when it announces this week a partnership with NuView Technology to use its ClusterX technology.
Despite the recent $US17 billion trade agreement between IBM and Dell to exchange technology and products, Dell is striking out on its own by partnering with NuView, a software start-up, to build on top of Microsoft's clustering technology. ClusterX offers load-balancing capabilities similar to Cornhusker, as well as security features in n-node clustering.
Unlike IBM, however, Dell will not scale beyond eight processors in a single box, according to Bob Van Steenberg, vice president and general manager of Dell's Enterprise Server division.
"We pretty much think eight way is the optimal answer to push the SMP envelope, even for Merced," Steenberg said. "To get beyond that we see clustering two [nodes], four [nodes], or eight [nodes] to get beyond the eight-way limit."
According to Illuminata's Eunice, the difference of opinion is really a tale of two companies.
"Dell is wrong. You can go beyond eight way. Dell is right in that it is a practical limit for completely off-the-shelf technology but Dell's whole world view is vanilla," Eunice said. "If you look at the big system players, they can do a lot better than vanilla. It will just cost you more."
IBM's Groves pointed out that with large SMP systems there is a single point of failure.
"If you have one copy of the OS running on 32 processors, it can take down the whole system," Groves said.
But according to Groves, the downside of clustering is the lack of applications written for clusters and managing clustered systems.