Vendors Show Off Four-Node Datacenter Configurations

REDMOND, Wash. - At least three OEMs are working on four-node failover clustering configurations for delivery soon after the release of Windows 2000 Datacenter Server.

Hewlett-Packard Co. (, Compaq Computer Corp. (, and IBM Corp. ( each demonstrated greater-than-two-node clustering capabilities at a Windows 2000 Datacenter Reviewer's Workshop here this month.

Four-node failover clustering is one of several high-end features Microsoft Corp. ( is adding to the Windows 2000 product line with Datacenter Server. The server product will also introduce to Windows support for up to 64 GB of RAM and 32-processor SMP support.

The type of clustering being addressed is Microsoft Cluster Service (MSCS) or Wolfpack clustering. Today, this clustering is available only in two-node configurations through Windows NT 4.0 Server, Enterprise Edition, or Windows 2000 Advanced Server. Machines configured as Wolfpack clusters share disk and monitor one another's health. As soon as one server in such a cluster notices that another has gone down, it takes over the applications or services that were running on the other machine.

HP and Compaq each demonstrated a four-node configuration for reviewers; IBM had to settle for a three-node cluster because a machine was lost in handling en route to the workshop. IBM officials say the company is on track to deliver a four-node configuration.

Microsoft instituted a rigorous certification process for Datacenter called the Windows Datacenter Program. The operating system will only be sold as part of a complete system that has been put through a detailed 14-day test. While most of about a dozen OEMs will offer single eight-processor machine configurations, those who want to offer more complex Datacenter configurations must certify them separately with Microsoft. Service Packs and upgrades will be administered through professional services.

HP and Compaq officials say they expect to have eight-processor machines in four-node configurations certified and going to customers once the operating system is approved for production.

The HP demo on four, eight-processor machines showed load balancing and failover for SQL Server and Exchange Server. An HP engineer showed how in failover situations, clustered machines can pick up the slack from failed machines, even reallocate duties between machines. As the Exchange server went down, a server set up for SQL duties took over the operation.

IBM demonstrated failover for financial services applications. Stock and mutual fund trades were each processed on one machine, while another machine handled bond and option transactions. An IBM representative killed the power to the node handling fund and option transactions, forcing a hard shutdown. He was easily able to get the application back online by distributing it to the other two nodes. A fourth machine was used to send application data to each of the server machines, which processed the transactions.

The four-node clustering demonstrations on eight-processor machines put to rest one two scalability rumors that have dogged Datacenter Server. One was that Microsoft and its partner Unisys Corp. ( were having problems getting one copy of Datacenter to scale across 32-processors on a single machine. Unisys squelched that rumor a few months ago with a 32-way demo on its ES7000 machine. The other rumor was that Microsoft was having trouble scaling four-node configurations on machines with more than four processors. This would have been problematic since the Windows Datacenter Program requires that the operating system not be sold on machines with less than eight-processor capacity.

A Compaq spokesman said that although there were issues in getting enough hardware for the certification team to test, there were no known technical issues in scaling.