Complexities Of Cluster Backups

• 08/30/1998

Microsoft's Clustering Service Raises New Questions

With the increasing deployment of Microsoft Windows NT Server Enterprise Edition,server clustering is becoming popular as a means to achieve high-availability without theexpense of deploying fault-tolerant redundant systems. The clustered approach differs fromredundancy in that, with clustering, secondary servers -- known as "nodes" in aclustered environment -- are in use performing their primary tasks, like providing e-mailor Web access services. Redundant systems simply mirror the functions of the primaryserver and perform no additional tasks. Clusters are classified as highly available withan average up-time of about 99 percent. Redundant systems are classified as fault tolerantwith 100 percent up-time under all but the most cataclysmic circumstances.

Given the high costs of redundant systems, many network architects are opting forMicrosoft's clustering architecture and its promise of near-continuous availability.Running on Windows NT Server, Microsoft's Clustering Service software provides theframework for determining how the cluster will handle the failure of any components.

IMPLICATIONS FOR DATA PROTECTION

While it may reduce the cost of building highly available systems, clustering magnifiesthe complexity of backup and recovery. In a clustered environment, the two standard backupmethods -- backing up over the network or to a directly attached device -- are eitherimpractical or inadequate when safeguarding all information in a cluster. While backup ofsmaller amounts of data can be conducted over the network, standard networks areinadequate for larger backups. To achieve acceptable performance on large data sets, anetwork administrator must back up the cluster's physical nodes to a direct-attached tapesystem. But this approach presents its own set of problems.

Clustered applications are packaged as "groups." Included in a group is thevirtual server name for the application, which lets a given application run on differentsystems at any given moment. To get a complete system image containing all the informationneeded to rebuild the cluster, each physical node must be backed up, not just the virtualnodes that represent the application groups.

Because clusters are inherently dynamic, responsibility for managing data andapplications moves from server to server as needed to keep the system running. Suchshifting of responsibility is called "failover" -- and failovers causeuncertainty in backups. A physical node's configuration can change depending on where thevirtual node resides. Therefore, an application could execute on one system at the time ofbackup and on another at the time of restoration. Unless the configuration of the physicalnodes during backup is identical to configuration during restoration, inconsistencies mayarise which make it impossible to guarantee that all critical data is recoverable.

Solutions for protecting clustered data are still evolving. By the year 2000, though,the puzzle pieces should be in place. For networks with smaller storage requirements,over-the-network backup is still the best option. Enterprise environments with largestorage requirements, however, should consider directly attached tape backup devices foreach node, since over-the-network backup in these sites consumes an unacceptablepercentage of bandwidth.

In this scenario, each server in the cluster has its own tape backup system.Administrators can manually back up each node in the cluster, but the backup applicationsdon't know that the node is part of a cluster. Directly attached tape backup devices areonly capable of taking a snapshot of the cluster at the time of backup; they don'tautomatically adapt to changes in cluster configuration caused by events like failovers.

Very soon, most software that runs tape backup systems should become"cluster-aware." This means that backup software understands the dynamics of theclustered environment and makes calls into the cluster to determine where data resides andhow to back it up. If restoration is needed, cluster-aware backup applications understandhow to interrogate the cluster to determine storage configuration and the proper order inwhich to restore the various elements to maintain consistency.

The next step is the intelligent use of direct-attached backup devices. Choosing amongthe pool of tape devices attached to the physical nodes in a cluster, the applicationdynamically specifies the proper tape device at time of backup and tells the device toback up designated files, directories or applications.

The tape backup application, because it has kept track of clustered activity, knowsexactly where the relevant data is located. This scenario provides automatically optimizedbackup, doesn't saturate the production network and frees IS staff from manualintervention in cluster backup operations.

Within the next year, the most cost-effective and efficient cluster backup solutionwill be achieved by building on prior developments -- cluster-aware backup software andintelligent direct-attached backup devices -- to allow the cluster to share a singlebackup device, like a tape library or tape autoloader. And while such a connection couldbe achieved over an existing technology like SCSI, future clustered servers are likely tobe attached to each other, and to shared storage and backup devices on a separate storagearea network, or SAN, using Fibre Channel technology.

With the advent of efficient, affordable SAN technologies and cluster-optimized backupcontrol software, network planners will finally be able to take advantage of the best ofboth worlds -- clustered servers backing up to a single, network-attached DLT or tapelibrary system.