In band or Out of band, it’s irrelevant
by Hu Yoshida on Nov 8, 2005
Almost everyone who talks about storage virtualization tries to classify virtualization as either “In Band” or “out of Band”. This classification of storage virtualization is only relevant to SAN based virtualization, if at all. Jon Toigo, an analyst who is known for his non BS approach to analysis says: “There is no such thing as in-band versus out-of-band: I can only assume that XXXXX (another analyst) was paid good money to create such a specious distinction.” Virtualize This! The general concept of “In Band” is that the mapping of virtual to physical is done in an appliance that sits in the data path between the host application and the target device. With “Out of Band”, the virtual to physical mapping is done in an appliance that sits out side the data path and the mapping is communicated over a control path to SAN switches (or Host Bus Adapters) that sit in the data path between the host application and the target device. The net result is the same, they both sit in the data path in order to intercept I/O and remap them to another storage address or addresses.
In band versus out of band has no relevance to the controller-based storage virtualization implemented in the TagmaStore Universal Storage Platform and Network Storage Controller. The difference is that the USP/NSC Control Unit is the target device for the host application, it is not a proxy sitting “in band” or “out of band” between the application and the storage. The Applications are talking directly to a device address in the USP/NSC control unit, issuing read, write, and other control and query commands. TagmaStore does not need to intercept I/O, and crack open the FC packets to see who the host application is talking to and what it wants done.
The storage virtualization that the USP/NSC does is essentially the same type of virtualization that all enterprise storage controllers do. It takes logical devices that are created by RAID striping across multiple physical disks and presents them in the USP/NSC cache to a host application as a Logical Unit Number or a mainframe volume type, depending on the host that the application is running on. The difference with USP/NSC is that these logical devices can be external as well as internal to the storage controller and can reside on other vendor’s storage units. The other major difference is that the USP controller is able to non-disruptively scale up to 32 petabytes of capacity (the NSC scales to 16PB) and tens of thousands of host connections. USP/NSC virtualization is just an evolutionary step in the desegregation of controller functionality from commodity disk arrays. All the controller functionality that Hitachi has developed over the years, like ShadowImage (in-system replication), TrueCopy (remote replication), Copy-on-Write Snapshot, etc. can be extended to heterogeneous externallyl attached storage systems.
Asking whether USP/NSC virtualization is “in band” or “out of band”, is a little like asking when did you stop beating your wife. The assumption is that you are doing virtualization like everybody else, in the storage area network. USP/NSC virtualization is not done in the network. It is a much simpler approach that transcends the Storage Area Network.
Comments (2 )
TagmaStore takes the central place among HDS products. It is very well suited for Enterprise market, but smaller companies can afford only modular storages like AMS/WMS/Thunder. What is your vision how to implement virtualization at SMB market? Should these companies back to host-based virtualization methods and utilize volume managers, file systems and host-based replications to make storage infrastructure non disruptive?
In response to Pavel Mitskevich’s comment, the TagmaStore USP control unit was scaled down for the midrange market and packaged in a 19 inch rack with 200V single phase power for the non data center environment. This product, the NSC55, has the same functionality as the USP but instead of 128 processors sharing a global cache, it has 32 processors sharing a global cache. It is priced for the midrange market. It provides storage virtualization at a lower cost than SAN based virtualization, when you look at the total cost of clustered virtualization appliances, additional switch ports, virtualization software, and the required storage. The architecture is such that it can scale to even smaller configurations in the future. For the SMB where there are only a few storage frames, it may be cost effective to use host based virtualization and replication products which have been tried and true for some time.
Virtualization and replication is best done at the end points of an I/O, either at the initiator (host) or at the target (storage Controller). There is no spoofing or remapping in between which can lead to additional latency and integrity/availability exposures. With a small number of initiators and targets, it is cost effective to do this in the host, but as the numbers increase it becomes more cost effective and more efficient to offload these functions to the storage controller, if that storage controller has the performance, scalability, and availability features of the TagmaStore USP and NSC.