Compaq Demolishes NT Cluster Benchmark

The latest Transaction Processing Council (TPC) benchmark test of a Windows NT cluster blows away previous results on Windows NT. The results of Compaq Computer Corp.'s configuration are pushing the platform into competition with more exotic systems from Sun Microsystems Inc. and IBM Corp.

Using six eight-way Pentium Xeon machines and the Parallel Server technology from Oracle Corp. (www.oracle.com), the Compaq cluster processed 99,274.90 transactions per minute (tpmC), shattering the previous mark of 50,208.43 tpmC for a Windows NT-based cluster.

The Compaq (www.compaq.com) system, the PDC/O2000, was configured with six Compaq ProLiant 8500 servers, each with Intel Corp.'s (www.intel.com) Pentium Xeon 550 MHz processors, 2 MB of cache and 4 GB of RAM. The servers were connected to 90 Compaq RAID Array 4000 disk subsystems through six Gadzoox Capellix 3000 Modular SAN switches, each configured with 22 ports. Total disk storage capacity was 9.8 TB.

The servers were running Microsoft Corp.'s (www.microsoft.com) Windows NT 4.0 Server, Enterprise Edition; Oracle Parallel Server; and the Oracle8i, Enterprise Edition, database manager version 8.1.6.

Mike Nikolaiev, TPC representative for Compaq, says several factors increased the throughput. Compaq's ServerNet technology, a mix of software and network cards, played a significant role. Also, the shared disk system was connected using Fibre Channel switches. "All six nodes could see all [9.8] TB of data," he explains.

These results are a boon for customers looking to increase performance in their clustered environments, Nikolaiev says. "There's a large subset of customers today running clustered environments to increase scalability," he says. "Not only can you add machines to increase throughput, but if a single node fails you can continue to run."

The price point Compaq reached -- $39.14 per tpmC -- is far lower than any other result in the TPC's top 10 transaction results by performance. The next closest is the No. 1 performance machine from IBM (www.ibm.com): The RISC System 6000 Enterprise Server S80 gets 135,815 tpmC at $52.70 per tpmC.

Sun Microsystems (www.sun.com) ranks third in the TPC top 10 performance chart. Its Enterprise 6500 Cluster pulling in 135,461 tpmC at a far more expensive price point of $97.10 per tpmC. "For a customer looking at this performance, NT will be the most cost effective way to get them there," Nikolaiev contends.

The previous TPC (www.tpc.org) benchmark lead on Windows NT was held by NEC USA Inc. (www.nec.com) when it recorded 50,208.43 tpmC, also using Oracle8i. It only used four eight-way machines rather than the six that Compaq used.

From September to December, Compaq, Dell Computer Corp. (www.dell.com), Fujitsu America (www.fujitsu.com), and Unisys Corp. (www.unisys.com) each released a Windows NT benchmark on a single eight-way machine, producing results from 40,168.5 to 40, 697.2 tpmC. Nikolaiev says the single node benchmarks are still important to test applications that don't scale.

It may seem that if 40,000 tpmC can be processed with one eight-way node, then six eight-way nodes should get results in the vicinity of 240,000. But Vince Gayman, director of marketing for industry standard clustering for Compaq, explains that Oracle Parallel Server has a lock manager that manages the access the six nodes have to the database so each node can be active at the same time. This presents significant overhead, but that overhead is considered acceptable because of the significant increase in availability and reliability.

Future of benchmarks, Nikolaiev says, will revolve around Windows 2000. The benchmark unit at Compaq is moving quickly to Windows 2000. Nikolaiev says he can't release specific results yet, but he can say there already has been vast improvements in performance. So much so that Windows NT benchmarks will probably be retired.

"I don't have any [Windows NT benchmarks] currently scheduled," he says. "I don't see why I'd want to go back when Windows 2000 is more scalable. Most applications are moving to Windows 2000 so we'll move our benchmarks there."

James Gruener, analyst at the Aberdeen Group (www.aberdeen.com), says benchmarks are great indicators of where the performance is, but real world tangible workloads are more important. "Research we've done shows these benchmarks are not taken nearly as seriously by the IT administrators as by the actual suppliers," Gruener says. "[Benchmarks] are a consideration but never the decision-making factor."

Gruener also explains that while these benchmarks don't necessarily have Sun and IBM quaking in their boots, the anticipation of performance improvements from future Windows 2000 benchmarks just may shift the landscape a bit.