Running mission-critical applications on an unproven operating system is too dicey for many corporations to stomach. But three trailblazers now running Windows 2000 Datacenter Server say they had good reasons to consider it.
Microsoft Corp.'s Windows Datacenter Program calls for well-established hardware makers to sell, support and test the product in configurations guaranteed to produce the sort of reliable and stable results that customers of rival Unix systems expect.
And Windows Datacenter Server is a logical consideration for companies using Microsoft software, particularly those pushing the limits of eight-way boxes running SQL Server on Windows 2000 Advanced Server. Datacenter Server scales to 32 processors, and its license allows for as many partitions as desired on a single physical server.
So far, the only choice for a 32-way box has been the ES7000 from Unisys Corp. in Blue Bell, Pa. But three early adopters who were glad to find the ES7000 say their Datacenter-based systems have largely delivered on the promise of high availability, scalability and reliability.
Mary Kay Consolidates SQL
Mary Kay Inc. in Addison, Texas, purchased a 32-way ES7000 server in April 2001 to cope with the explosive growth of its online offerings.
Prior to that, the cosmetics, skin care and dietary products manufacturer and distributor had clustered four-way servers, upgraded processors and shifted from Windows NT to Windows 2000 Advanced Server.
But Morris Koeneke, database services manager, says distributing databases evenly was becoming difficult. Database replication was gobbling up resources. Single-user interactions could trigger requests to multiple database servers, slowing the network. And Mary Kay was outgrowing servers before its leases were up.
The company went live in June 2001 with its ES7000, initially on Windows Advanced Server. Using Unisys tools, IT staffers carved up the 32-processor, 32GB frame into four eight-CPU partitions and configured them as two clusters: one for SQL Server 2000 production databases and the other for testing, staging and fail-over.
Mary Kay was able to consolidate four production database servers to two, but it soon needed more horsepower and shifted to Windows Datacenter in April 2002. Administrators converted the four eight-way partitions to two 16-CPU, 16GB partitions: one for the production SQL Server and the other for backup.
"Datacenter is basically Advanced Server on steroids," Koeneke says. "It's the same operating system. It just has gates opened up that let you manage large servers."
Consolidating to a single SQL Server instance for e-commerce improved performance and eliminated redundant databases and replication headaches, says Koeneke. Processing capacity went from 1,800 orders per hour to more than 3,000. And Koeneke says problems were scarce and insignificant compared with prior Windows versions.
Mary Kay purchased three more ES7000s last year for a supply chain initiative. One has a 16-CPU instance of SQL Server and two eight-CPU partitions of BizTalk Server. Another has SQL Server running on 12 CPUs, clustered to the 16-CPU SQL Server on the other box, and five four-way partitions for Mary Kay's J.D. Edwards & Co. enterprise resource planning (ERP) system. The third box is for development and testing.
Plans call for the BizTalk and ERP database servers to be consolidated. Koeneke says he also hopes to run a four-way Datacenter cluster with three active 16-way nodes and one passive 16-way node hosting all of the company's databases.
Koeneke says he isn't convinced that total cost of ownership is cheaper with Datacenter Server than it would be with Unix, but given Mary Kay's commitment to Microsoft products, Datacenter made sense. His only gripe is with the Datacenter program, which Microsoft is revamping. Under the existing program, the hardware provider is supposed to be the single point of contact for problems, yet Koeneke says he often waits for answers as vendors point fingers.
Koeneke would also like more choice in the subsystems that attach to his Datacenter Server beyond those in the exact configuration Unisys tested. "They tell us how great everyone gets along and that we have one throat to choke," he says. "But the reality is we get total certification on a solution provided by one OEM. Our flexibility is limited."
Premera Downsizes
 |  |
| Premera CIO Alan Smit | |
CIO Alan Smit says he saw an opportunity to focus on a single vendor, even though he shared the general industry skepticism about Microsoft's ability to support enterprise transaction processing.
Careful analysis of the technology, benchmarks and Datacenter program helped ease his concerns. So did the Windows-based client/server claims and membership administration system Premera bought from TriZetto Group Inc. in Newport Beach, Calif.
In July 2001, Premera went live with two ES7000s running Datacenter Server -- one 32-way system and one 16-way box in a clustered, fault-tolerant configuration. IT staffers created an eight-way partition on the 32-way box to run SQL Server and used a clustered eight-way partition on the 16-way box as a fail-over system. That paid off late last year when a memory problem caused the active server to crash. Users noticed only a 20-second slowdown as the second node took over the workload, says Bob Crownhart, an IT director at Premera.
Premera now has 172,000 members loaded into the Datacenter-based system and has been running nonstop since August, when it took down the system to address a hardware problem, says Crownhart.
The company plans to move to 16-way partitions as it scales to more than 1 million members. "Knowing we can scale beyond eight-way partitions is the biggest comfort. That's what we needed to convince ourselves that we could scale to 1.4 million members and beyond," Crownhart says, noting that the target completion date is 2005.
Premera is also looking to shift online analytic processing from its mainframes to Datacenter Server. "The fewer platforms I have to support, the better," Smit says.
First American Maxes Out
|  |
| First American CIO Larry Godec |
At the time, the Santa Ana, Calif.-based insurer was halfway through the rollout of its homegrown First American Software Technology (FAST) Transaction System, a consolidated title and escrow system for 10,000 users in 1,200 offices, says CIO Larry Godec. Tuning the application to make fewer database calls was one way to solve the problem. Porting the database server to Unix was another possibility, but that would have meant an extra six months to rewrite application code in Java, says Godec.
Yet another choice was splitting the database and creating a number of regional databases. However, Godec frowned on that option from a support and maintainability standpoint.
First American was adding hundreds of users each week and didn't want to stop the rollout's momentum, so it turned to a bigger box for its application's SQL Server 2000.
Godec figured that switching to an ES7000 wouldn't stop the rollout for more than a few weeks. Coordinating the setup with Unisys, Microsoft and EMC Corp., which was supplying a storage-area network, required about six weeks. The conversion itself took 12 to 16 hours.
But the rush through the migration process wasn't without incident. A bad ES7000 CPU caused a business interruption, and incorrect settings caused two more outages, Godec says. He advises users to make sure all processors are burned in appropriately and all settings are triple-checked before going live.
As the system load soared, First American encountered a problem with the Datacenter Server's scheduler, which is part of the operating system's kernel. The operating system froze when the scheduler backed up under extremely heavy load, Godec says.
"Adding that number of users that quickly to a system, you're going to find the stress points pretty quick," he says.
Microsoft tested the application to identify the problem and developed a more efficient scheduler that will ship in April with Windows Server 2003, Godec says. First American plans to upgrade as soon as it can.
In the meantime, turning off the affinity setting, which can dedicate processors to a specific system, such as the operating system or the database, has helped to offload some of the work from SQL Server and reduce the overall load, Godec says. Reducing the frequency of the database monitoring and replication on the EMC storage system also helped, he adds.
Unlike other companies that partition their ES7000s, First American has all 32 processors managing the workload of the operating system and the database. Godec says he considered 16-way partitions, but uncertainty about the effect on utilization led his team to use all the processors and add a second ES7000 at its Dallas site for redundancy and fault tolerance.
The Unisys box now runs at 40% capacity, so First American has plenty of headroom, Godec says. And he adds that he has been happy with the Datacenter Server technology.
"I would equate the ES7000 to other things that I've known and run in the past, which include the E10000 from Sun, and the Hitachi and IBM mainframes," he says. "It's every bit as reliable."
Godec says he never had a scheduler problem on an IBM mainframe, but he has encountered problems on large systems where the vendor response has been, "You need to upgrade to the new release."