While clusters that incorporate both failover and scalability are available in the minicomputer and high-end RISC-based Unix market, managers have been looking for clustering solutions on Intel-based systems, hoping to reduce the costs associated with the high-end systems and to add the benefits of clustering to existing applications running on NetWare, Windows, and Unix.

Network operating system vendors such as Novell, Microsoft, and SCO have responded, albeit slowly. Failover functionality has been available on Intel platforms for quite a while, although typically in a two-node system. Each node must be able to take over all the applications running on the other, which means that each can be loaded to about only 50% of its capacity.

Scalability hasn't been available at all until very recently, and it's still quite limited in functionality compared with the larger systems. Larger systems have a single system image, in which all the nodes in a cluster share their memory and other resources, and an application can run across all nodes at once.

Intel-based clusters offer scalability based on two models: threading and IP load balancing. Threading divides an application into different threads of execution, which can be executed on different processors and on different nodes. Typical applications that work well with threading are graphics rendering, mathematical modeling, and others that perform multiple separate operations on pieces of data.

IP load balancing runs multiple versions of an application and distributes requests among the different versions, according to which server is least loaded. Applications that typically work well with load balancing are ones with relatively static content that must be made available to multiple users, such as Web servers, and applications that route data, such as firewalls, encryption or virtual private network routers, and File Transfer Protocol servers.

There are many load-balancing products that work well and don't require integration with the network operating system, such as hardware-based switches and network appliances.

There are applications that can't be scaled by threading or load balancing. A database, for instance, tends to be a single, massive system that can't easily be threaded. A database requires a single large chunk of memory and works with dynamic data structures, which load balancers can't handle well. Ironically, managers want to cluster database servers for performance optimization.

There are proprietary databases that scale to four or more nodes from IBM and Oracle on Windows NT Server, but they're all closed solutions. Migrating from one database to another is extremely difficult.

I looked at three network operating system clustering offerings recently: SCO UnixWare 7 NonStop Clusters, NetWare 5.1 Cluster Services, and Windows 2000 Advanced Server with Cluster Services and Network Load Balancing.

Most managers will probably make a decision to implement these clustering services based on their existing network operating system. All are reasonably simple for a trained administrator to install, but nothing for the novice to attempt. The type of application already in place will also help determine the choice of product. The desire to scale an existing server rather than port the application to a new, faster platform is one of the basic reasons for clustering.

SCO's UnixWare offers functionality that is most similar to the clustering found on high-end systems. It includes some scalability and unique management features that make for easy installation of additional cluster nodes, as long as the administrators are familiar with Unix. NetWare offers administration that is fully integrated with the Novell Directory Services tools, and it provides four-node failover but not application scalability. Windows 2000 Advanced Server offers a two-node failover solution and provides IP load balancing.

UnixWare allows clustering of up to six nodes on standard Intel servers, with functionality much closer to proprietary Unix solutions than the other products supporting standard UnixWare applications.