UnixWare uses two channels for intracluster communication: ServerNet for interserver communications and Fibre Channel for storage. With redundant ServerNet switches and redundant Fibre Channel connections, it's possible to swap out every component of the cluster without bringing the cluster down.

Once the primary node is installed, configuring the rest of the nodes is a snap. The installer program generates a two-disk set that is used to create the rest of the nodes. Insert the first disk in the second node, then the second when prompted; tell it the node number; enter some basic network configuration and license information; reboot the system; and it joins the cluster.

Unlike the other clusters, the operating system is contained on the shared storage system. The only thing that needs to be on the local drive is a boot kernel that tells the system how to find cluster resources. All resources on all the systems in the cluster, including local disks, are available to any node in the cluster. This makes installing cluster nodes and applica-tions incredibly simple. The management tools are available from the graphical user interface or from the command line and are straightforward and relatively simple for Unix.

Applications don't need to be installed more than once, no matter how many nodes they will run on. Management of the cluster is via the GUI from the primary node or browser, command-line interface from any node, or a telnet session.

Failover from one node to another is generally a matter of less than a minute, although that depends to a certain extent on the applications installed. It's also straightforward to configure. In normal operation, processes are shared among nodes automatically, dynamically balancing the load across the cluster. Adding additional nodes adds about 95% of the capacity of the node. UnixWare applications can run unmodified (cluster unaware) and can be optimized to exploit cluster features.

The limitation is that applications must be able to run on a single node or multiple instances on separate nodes. Databases can't run as a single application on multiple nodes. So Web and E-mail servers and similar applications can easily scale across up to six nodes, but databases will still require specialized proprietary clustering software, at least for now.

NetWare Cluster Services for NetWare 5 allows up to eight nodes in a cluster out of the box and will support up to 32 nodes with special support from Novell. This allows servers to be more efficiently utilized. Individual servers can carry heavier loads, since each app can be distributed to a different server, rather than having only one other server to fail over to.

NetWare supports failover clustering for both applications and network services, including file and print, dynamic host configuration, and domain name service. This allows administrators to cluster critical network services as well as applications such as Web or database servers.

NetWare is installed on two or more servers in what is to be the cluster. The servers share a storage device, either a Fibre Channel storage system shared through a daisy chain or Fibre Channel hub or switch, or a shared SCSI device. The shared volume allows files from one server to be accessed by the server that is taking over its duties. The servers communicate synchronization information over the standard network interface. A separate network to hook the servers together is not required, although it can be set up that way.

NetWare was the only one of the three clusters that I installed from scratch--both SCO and Microsoft sent engineers. I installed it on four Gateway 7300 servers, using a shared SCSI disk array. The most difficult part of the configuration was installing NetWare on all four servers. If administrators are contemplating a large cluster, they would be well-advised to look into a software-distribution mechanism, such as Novell's ZENworks for Servers. The installation is straightforward and relatively simple, considering the overall complexity of the system.

In general, enabling applications is a simple process. The application is installed on the primary server that will run it and may need to be installed completely on the other nodes it will need to fail over to. Then an application object is created for the cluster using ConsoleOne from the administrator's console.

The object contains information about the application, such as the IP address that will be used to access the application and the scripts that will run on the server that fails and on the new server the application moves to. It also contains information on the designated cluster nodes that the application will move to, how it should be moved (automatically or manually), and when it should move back to the original server (fail back). These objects are administered through the same ConsoleOne/NWAdmin interface as any other NetWare objects--a familiar task to any NetWare administrator.