A Tale of Scale: How Wintel Rode Profusion to the Glass House
- By Scott Bekker
For a long time it was safe to write off Wintel as a scalable platform.
But, for the last three years, a combination of factors have sent Microsoft Corp. and Intel Corp. careening toward the glass house at a rate that should give UNIX/RISC vendors pause and give IT executives reason to take notice of that upstart PC platform.
The lifecycle is winding down for a central player in Wintel scalability: Intel’s Profusion chipset. The eight-processor Profusion board provided a launching platform for nearly two years of rapid Wintel scalability gains.
Even as Profusion-based systems are being ushered out of the spotlight, new server system technologies are queuing up to take Windows and Intel more squarely into raw performance competition with the UNIX offerings from Sun Microsystems Inc., IBM Corp. and HP.
And, as the raw performance improves, Wintel is maintaining its price-performance edge over traditional UNIX systems.
In 1998, Windows NT was perceived as good only for four-processor servers. Benchmarks published with the Transaction Processing Performance Council (TPC), the most closely watched scalability benchmark, bore out the perception. Three eight-processor systems tested that year yielded anemic increases in scalability of 34 percent to 42 percent better than their four-processor counterparts on the TPC-C benchmark, a test of OLTP scalability.
Opinion was mixed on whether it was the operating system; Microsoft’s flagship database, SQL Server 6.5; or the hardware that was to blame for the poor results.
Early Signs Wintel Might Someday Be Credible
In 1999, a few positive signs emerged that Microsoft and Intel might be serious about creating enterprise-class systems with Windows NT on Intel. First, Microsoft began shipping SQL Server 7.0, a completely overhauled database.
Then Microsoft released its fifth Service Pack for Windows NT 4.0, a rollup of bug fixes that, unlike some previous service packs, fixed more problems than it created.
Intel delivered the Profusion chipset that it had been promising since acquiring the technology in 1997. At the same time, Intel shipped faster and faster processors, including 550 MHz Pentium III Xeons, a big improvement over the 200 MHz Pentium Pro processors the earlier eight-processor system benchmarks ran on.
Unisys Corp. ran the first Windows NT 4.0/SQL Server 7.0 benchmark on an eight-processor Profusion system, and the results showed a 55 percent performance improvement over the best comparable four-processor Windows system of the time.
The result caused Giga Information Group Senior Analyst Brad Day to enthuse: "What’s really pretty exciting about this is that this is really the first time that a Wintel solution has been able to cross the threshold into the UNIX/RISC environment."
The usual finger-pointing about whether Microsoft scalability limitations rested with the operating system, the database or the hardware was no longer necessary. The Wintel architects had a chipset they could build on.
The 2000 Generation
Expected since 1998, Microsoft finally delivered the upgrade to Windows NT 4.0 in February 2000. Early adopters of Windows 2000 reported the same levels of reliability and stability that users found with Windows NT 4.0 Service Pack 5.
Intel cranked out faster processors, including 700 MHz Pentium III Xeons for four- and eight-processor servers.
Meanwhile, Microsoft readied two pieces of software heavily focused on the data center. One was Windows 2000 Datacenter Server, an operating system tuned for big systems with support for up to 32 processors and 64 GB of RAM.
The other product was a database robust enough to support Datacenter Server – SQL Server 2000. Both products launched in late September.
All through the year, OEMs ran TPC-C tests of Profusion-based eight-processor systems running Windows 2000 and SQL Server 2000. Dell, Fujitsu-Siemens, Compaq and IBM showed steadily improving scores.
The final TPC-C score of the year, posted by Unisys, was a 62 percent improvement over the first Profusion-based benchmark. The linear scalability versus four-way systems had improved, as well. The system performed 72 percent better than a similarly configured four-processor server.
Profusion Reaching Its Limit
This March, Intel released the final Pentium III Xeon processor: the 900 MHz.
Intel’s position is that the 900 MHz processor, with its 29 percent jump in clock speed over the 700 MHz, will bring significant performance increases.
"Obviously, the large frequency jump allows for a very large performance boost," says Shannon Poulin, Launch and Disclosure Manager in Intel’s enterprise platform marketing organization.
Jerry Buggert, who runs Unisys’ system analysis modeling and measurement organization and is chairman of the TPC, doesn’t expect quite so much out of the chip.
"Even putting the 900 MHz Xeons into the Profusion chipset is a real shoehorn activity. You’re becoming very limited by the memory throughput at that point," Buggert says.
Whether the chipset has any benchmark juice left in it or not, Buggert and Poulin agree it was monumental for the Wintel platform.
"It was a very, very strategic product for us. It enabled Intel to penetrate the back-end space and offer the price and performance leadership over some of the competitive products that were in that space," Poulin says.
Assessing Profusion Against UNIX
Microsoft and Intel definitely proved the platform against itself, delivering real scalability through Profusion over the platform’s traditional strong point in four-processor systems.
On price-performance, the systems also track closely to the numbers posted by Wintel four-processor systems, trouncing UNIX/RISC systems.
Comparing Profusion-based systems performance against the raw performance of the best UNIX systems, however, shows the Wintel alliance lost ground over the life of the chipset and gained very little with Profusion compared to where Microsoft and Intel were with pre-Profusion systems.
When Data General posted the first Wintel TPC-C eight-processor benchmark in 1998, the result was 31 percent of the performance of the top UNIX system at the time, a 16-processor HP 9000 running HP-UX and a Sybase database.
When Unisys posted the most recent Wintel TPC-C eight-processor benchmark in November, the result was 28 percent of the performance of the top UNIX system at the time, a 24-processor IBM AIX/Oracle8i system.
On the Verge
A number of hardware developments may break the raw performance standoff soon. During the most explosive period of Profusion system benchmarking, Unisys began shipping its e-action ES7000 server. The 32-processor servers are built on a new architecture developed at Unisys called Cellular MultiProcessing (CMP).
Designed specifically with Windows 2000 Datacenter Server in mind, the systems use standard Intel Pentium III Xeon processors.
Unisys officials say they are working on a TPC-C benchmark.
On a quasi-industry standard benchmark of mySAP.com Sales & Distribution users, Unisys tested CMP systems with 16, 24 and 32 processors and published the results. The 24-processor system, running Datacenter Server, gave nearly twice the performance of the top-performing Profusion-based system. The 32-processor system eclipsed all but a handful of Unix results. The headroom that remains to Unisys gives the company’s engineers an opportunity to make a run at those top UNIX-based results.
Opportunities for the Wintel platform to scale exist on several fronts. There are Intel’s next generation of 32-bit Xeon processors, Intel’s 64-bit processors, Microsoft’s scale-out initiatives and other big hardware systems based on Intel processors that will challenge Unisys’ CMP.
Intel’s next 32-bit SMP processor, called Intel Xeon, stems from the faster microarchitecture it shares with the Pentium IV processors.
Unisys plans to modify its existing line of CMP servers to support the new line of Intel Xeons, making further 32-bit scalability gains possible on that architecture.
More possibilities open when considering Intel's recently delivered 64-bit Itanium processors.
Unisys should be ready to ship 64-bit systems shortly. NEC had a 16-way system called the AzusA awaiting Intel’s silicon, too.
IBM also revealed a new Intel processor-based server that will scale up to 64 processors, called the IBM eServer xSeries 430 in March. The NUMA-based system won’t support Windows until the next release of Windows Datacenter Server.
Another possibility is the scale-out route to large systems that Microsoft promotes heavily. Microsoft, IBM and Compaq all promote benchmark results that show Windows 2000 at the top of the TPC-C scalability list with results as high as three times greater than the nearest UNIX/RISC vendor’s score.
The results depend on a promising method of clustering Windows servers together for performance largely unproven in real environments.
The enterprise administrator’s checklist for data-center class systems is simple, but demanding. Systems must be rock-solid from applications to operating system to underlying hardware. Support must be immediate and accountable. Headroom must exist to allow applications to grow.
Microsoft and its partners have worked furiously to check off those boxes. They’ve posted checkmarks in stability and support, and they delivered improvements in scalability with the Profusion chipset.
Beyond Profusion, they’re delivering a little more headroom every week through scalability enhancements.
For the very biggest new applications, UNIX/RISC systems retain an immediate headroom advantage that makes them the logical choice.
Where Wintel’s compelling price-performance and improving raw performance make it worth a hard look is for new data-center applications with the potential to become very large.
HP, Compaq Back Away from Wintel Server Plans
64-bit Windows Server Sales to Come Through OEMs