This article can also be found in the Premium Editorial Download "Storage magazine: Is storage virtualization ready for the masses?."
Download it now to read this article plus other related content.
Sophisticated roll-back capabilities are being developed to instantly roll back a complete volume. Currently, a snapshot is often reset on a routine basis or disabled upon completion of the backup to minimize the capacity and performance impact. However, with sophisticated roll back and efficient checkpointing, soon users will be able to select the time of the restore point and instantly revert to that moment.
Note that with copy-on-write, two writes now occur when production data gets updated. First, the original value must be saved in the snapshot index, and then the change is made to the production base volume. There are variations of the copy-on-write that don't require the original value to be saved in a different location, however, compacting the disk will be required with these variations. Such variations include the redirect-on-write and log-structured file methods.
Copy-on-write can be implemented with server-based software in the storage subsystem, and more recently in intelligent switch or virtualization devices. Each has pros and cons associated with it. (see "Copy-on-write snapshots: pros and cons")
Disk mirroring has long been used to maintain two or more up-to-date full copies of the data. Every write request to the original data is automatically duplicated to other mirrors or copies of that data. The mirror may be contained in the same subsystem or be between different subsystems, although
The primary purpose for a mirror is disaster recovery. In the event the entire subsystem should fail, mirrors must be written between two subsystems and have the appropriate distance for the disaster to not affect both systems. Often two subsystems will be mirrored and sit in the same data center. This will guard against hardware failure, but not from a site catastrophe such as a tornado. The further the distance, the more delay in performance; thus, asynchronous modes of data transfer are available to accommodate wide-area distances. The cost must be weighed with the appropriate business risk.
A mirror provides real-time redundancy, and when it's active, isn't a frozen image or snapshot. The mirror can be temporarily suspended - also referred to as a broken or split-mirror - to create a snapshot or point-in-time copy. The disk subsystem is told to temporarily stop making updates to the mirrored copy so the data is frozen at the point of the suspension. The split-mirror can then be used for the backup process or other purposes.
Mirrors create an instant copy, or snapshot, of the data with the split capability. Unlike copy-on-write, a full data copy is available. In order to keep the disaster recovery copy available, a third mirror is usually established for the purpose of splitting. This requires three entire copies of the data volume to provide the protection and meet continuous processing for backup and other development needs. In this setup, there is a primary and secondary real-time copy, and a tertiary point-in-time copy of the data. Since the split-mirror or snapshot is an entire set of the data, the data can be updated for development or training purposes. In contrast to copy-on-write products, only some of the products allow writes to occur to the snapshot copy. Note that once the split-mirror has been modified, the entire mirror must be rewritten to establish an active mirror with the original volume.
When the backup is complete, the mirror is resumed. In more sophisticated systems, the writes have been saved from the point of the suspension and they are applied to the mirror and normal mirror operations resume. However, in some products, the entire mirror must be rewritten once the mirror is broken. In either case, the original data volume is not affected by breaking the mirror.
Products that utilize the split-mirror to provide an instant copy or snapshot, include:
- EMC TimeFinder (Symmetrix)
- Hitachi InstantSplit for ShadowImage (Lightning series)
- HP SureStore Business Copy (XP)
- Sun StorEdge Instant Image (9900 systems)
- Xiotech REDI (Magnitude)
- LSI Logic ContinuStor Director
- Veritas Volume Manager and Volume Replicator
Mirrors have a write penalty while the mirror is active. The write penalty ceases once the mirror is split, or when the snapshot is created. In contrast to copy-on-write, the write penalty begins after the snapshot is taken. In either case, the snapshot is available in an instant and doesn't require the user to wait for the copy to get created.
When to use copy-on-write vs. split-mirror
Copy-on-write is an efficient method of creating an image. Since it utilizes much less disk space, multiple copies can be kept as restore points or for other purposes, such as testing and training. Continuous development based on this technology will provide simpler and closer to the problem recovery points. Eventually, the user will select the source restore time and a file or volume will be available from that point-in-time. Although disk space is becoming much more economical and scalable, replicating large data sets is still expensive. Copy-on-write snapshots minimize disk space considerations, saving time and data with multiple roll-back copies. Some copy-on-write implementations don't allow the snapshot to be used as a source image that can be written to, so these implementations are limited to backup and recovery purposes.
Split-mirror leverages mirroring technology that's increasingly being deployed for disaster recovery purposes. Mirroring can be costly, but may be appropriate protection from disk subsystem failures and disasters. The ability to split the mirror in environments already utilizing mirroring is a natural extension. Another option is to take a copy-on-write snapshot of an active mirror in order to save on disk space and allow more point-in-time images. The split-mirror is effective for rapidly churning data, since there isn't a write penalty after the mirror is split. Since the write penalty is already completed prior to the snapshot, the split-mirror is effective for copies that will be utilized for a long time - for example, an extensive data mining deployment.
This was first published in June 2002