Server clustering gaining new clout

Electronic News, Oct 14, 1996 by Elaine Chen

New York--Clustering was poised to enter the mainstream at last week's Unix Expo, with Sun Microsystems debuting complete solutions from under $100,000 and other vendors claiming systems costing much less. However, analysts noted that the traditionally proprietary industry could be facing some major shakeups, with new technologies and Windows NT-based standards on the horizon.

Clustering, or linking separate computers to act as one, has long been associated with mainframe-class applications due to cost and complexity. However, as the technology has matured, clustering has become an attractive solution for a range of new customers. Its advantages include failover capability, high performance and scalability. As customers can often take advantage of legacy equipment while building more powerful systems via clustering, it may also prove a cost-friendly migration path for smaller users planning an expansive future.

Sun's new products bring clustering to the Ultra Enterprise servers announced this spring. (EN, April 29). Sun is currently offering turnkey two-node solutions in what it terms a "smart cluster" approach, with systems priced from under $100,000 to over $1 million and scaling from one to 30 processors. Special features include a new high-availability version of BEA Systems' Tuxedo transaction monitor, which is exclusive to Sun for the next six months. Future plans for the clustering line include up to 8-node capacity by next year, with integrated disaster recovery features, while 16-node capability is planned for 1998.

Meanwhile, IBM is offering clustering on a global scale with its High-Availability Clustering Multi-Processing, now featuring High Availability Geographic Clustering (HAGEO). By permitting data mirroring over a WAN, the HAGEO systems allows users to remain operational even in the event of a major disaster such as an earthquake. IBM clusters also include its "Phoenix" technology, which provides enhanced failure monitoring and recovery coordination capabilities as well as an open set of APIs designed to encourage application development across a variety of operating systems.

The open systems approach is a radical change for the traditionally proprietary realm of clustering. As stated by Jerry Sheridan, a principal analyst with DataQuest, there currently is "no industry standard for high availability," or, for that matter, even performance benchmarks. However, change may be coming soon, thanks to Microsoft's previously announced Wolfpack clustering specifications (EN, May 13). Another standards-based technology which could help transform the industry is NUMA, or non-uniform memory architecture, which is designed to improve the performance of clustered systems via a distributed, shared memory architecture.

Unisys is hoping to ride the open systems wave through its ClearPath Intel-based server clusters. According to Unisys product manager Andrew Dyamoke, Unisys offers up to four-node clusters of the SMP 6400 server line, which supports either NT or Unix and scales up to 10 CPUs. Unisys is eagerly awaiting Wolfpack and advanced technology for the NT platform. While Mr. Dyamoke admitted that NT is currently "not a scalable system (as it) doesn't have the resiliency," he stated, "There's no doubt in my mind that it will be there."

Unisys is looking to NUMA as well, thanks to a relationship with key NUMA player Data General (DG). DG VP Philip Gerskovich emphasized the performance capabilities NUMA could bring to Intel platforms, as it "allows you to use SMP programming, but take it to a larger system." DG hopes to take advantage of Intel's huge volumes by using NUMA to tie together lower-cost, standardized four-processor Intel motherboards. "The RISC guys are going to have a hard time competing on a price/performance basis," he added.

DG brought both NT and Unix systems to the show, with its Unix systems featuring 4-node clusters running DG/UX, and its two-node NT clusters using Firstwatch failover software from Veritas. Like Mr. Dyamoke of Unisys, Mr. Gerskovich agreed that while NT "doesn't scale up to the demands of the large enterprise...it will eventually." Accordingly, while NT systems and clusters may be relegated to the low end for now, in the long-term Mr. Gerskovich said the RISC players may see "Intel take over their whole operation."

Industry veterans Hewlett-Packard and Digital Equipment may provide some clues to the future confronting the high-end RISC companies. Both have thrown their lot in with the "Wintel" alliance, albeit from different approaches: HP will migrate to the future Intel Merced chip for even its technical computing needs, while Digital has wholeheartedly embraced NT on its Alpha RISC platform.

The HP endorsement means far more than just chip sales to Intel; by providing a migration path clear up to the supercomputing class, it could be just what Intel needs to break out of the low-end market. "It's an incredibly exciting arena we're entering," said HP high availability systems product manager Tad Walsh. HP offers eight-node clusters for its HP9000 server line, with administration via its HP ClusterView management platform. Mr. Walsh also said in the immediate future, HP expects to make some major announcements--and gain significant market share--in the disaster recovery market.