Storage area network (SAN) and network attached storage (NAS) technologies are moving to new distributed-storage architectures. Both are being reconfigured to combine network technologies like Ethernet and the Internet Protocol (IP) as a transport mechanism with storage protocols like SCSI and FCP to serve the needs of widely distributed applications.
NAS architectures attach small, relatively cheap devices called filers or thin servers or NAS appliances to a disk subsystem and then directly attach the combined subsystem onto the LAN. This makes it possible for one or more clients on the LAN to have direct access to the data contained in a NAS appliance without having to involve and interrupt an application server.
In general, NAS devices are connected to an Ethernet network and can scale storage without having an impact on other applications within the enterprise. Unlike classic client/server architectures, which require separate copies of files for Windows NT and UNIX, NAS filers operate independent of the client's O/S, making it possible for users to share files across multiple platforms.
While deployment of NAS has significant advantages within the enterprise, it does not fully solve all of the storage requirements. Some of the NAS benefits begin to diminish as enterprise storage requirements grow, driving increased management complexity; decreased performance due to shared LAN bandwidth limitation, and larger and larger file I/O sizes. For example, in high performance On Line Transaction Processing (OLTP), memory accesses are typically for small (4kbytes or smaller) random blocks (not complete files) within a very large database. The network overhead, coupled with the file system, would limit performance in this type of application.
Because of some fundamental issues with NAS, SANs have gained tremendous traction over the recent past. Fundamentally, what a SAN does is to place a dedicated storage network in between servers and storage subsystems. Unlike a DAS configuration, with an inflexible bus based topology between the server and storage, in a SAN, any server can connect to any storage subsystem servers sit in the middle, connected to client systems, and the outside world, via the datacom LAN. On the backside of the servers sits the SAN, now a full fabric based topology for interconnects, servers and storage devices.
Among the primary advantages that SANs have over DAS architectures is the any-to-any connectivity, for example, a storage subsystem is not completely dedicated to only one server. Also, SANs offer high availability features, utilizing redundant paths within the fabric to offer fail-over. For example, if Server A fails, the data stored on Subsystem C can still be accessed by Server B, all of this is done without ever involving the LAN.
In addition, SAN implementations can provide a lowered overall cost of storage management and increased data transfer rates. Other benefits include reduced LAN congestion, and backup procedures that do not involve the servers or the LAN. Scalability is also simplified, as SANs can grow and adapt much more readily than any bus based topology, as capacity can be readily added or re-allocated to where it is needed.
While Fibre Channel has achieved first mover status in terms of SAN protocols, other protocols such as iSCSI, iFCP and FCIP are also being discussed as part of an IP Storage design scheme. In the current approach to building a SAN, a Fibre Channel cable-of either copper or optical media-is used to connect a variety of servers and storage devices.
One of the major advantages a Fibre Channel SAN provides is the elimination of the physical distance limitation. For example, in a DAS implementation using Ultra 160 SCSI, the storage subsystem can be located no more than 25 meters (12 meters for a multi-drop implementation) away from the server. In a Fibre Channel SAN, the storage devices can be as far away as 10 kilometers or through the use of channel extender, as far as 200 kilometers. This makes it possible for large, centralized storage subsystems to support remote mirroring without requiring the use of the datacom LAN.
In the early going, Fibre Channel has taken the stage as the be-all end-all of SAN protocols. Interestingly, however, Fibre Channel itself is based on the SCSI protocol, with the fundamental change being that Fibre Channel changed SCSI from a parallel bus-based architecture to a serial interface fabric topology.
By adopting a serial interface and fabric topology, Fibre Channel, is able to overcome the distance limitation of a parallel interface. In addition to the greater distance, Fibre Channel also supports full duplex: It can send and receive data at the same time, effectively doubling its throughput. Utilizing Fibre Channel's 24-bit port address field (N Port ID), it overcomes the 16- node limitation for a SCSI bus implementation enabling large-scale systems to be implemented, centralized management capability, low latency and increased reliability using redundant paths.
Before crowning Fibre Channel as the presumptive winner for all SAN implementations, it's important to take a look at several versions of IP Storage currently under review by the Internet Engineering Task Force (IETF), a standards body for Internet architectures iSCSI and FCIP.
Since it is based on an existing, mature network protocol (TCP/IP) and has the potential to leverage the existing infrastructure, IP storage promises to have a significantly lower cost of ownership. For one thing, because of its larger installed base, more money is being spent on advancing Ethernet.
Beyond that, IP storage eliminates the need to have two networks (Ethernet for the servers and Fibre Channel for the SAN). In an IP SAN, one super network interface card (sNIC) can be used to access both the Ethernet and the IP SAN, compared to two cards (a NIC and a host bus adapter) to access Ethernet and the Fibre Channel Network, thereby saving money. Also, IP Storage is easier to manage since it eliminates the need to have network administrators learn multiple ways to control data and storage networking.
The iSCSI protocol sends SCSI commands as data (no different from Fibre Channel) but uses TCP/IP layers instead of Fibre Channel layers when it transports the data. In this context, iSCSI, running over TCP/IP and the link layer, replaces the FC2-4 layers that Fibre Channel uses.
The iFCP protocol also sends SCSI commands as data and uses TCP/IP for transport. The difference between iSCSI and iFCP is that iFCP keeps the FC-4 layer instead of replacing it with iSCSI. By using the FC-4 layer, iFCP can be used with current software applications that manage Fibre Channel networks and can also be used on both IP and Fibre Channel-based SANs.
In addition, FCIP is an open, standard method of tunneling. FCIP takes a Fibre Channel frame of information and then encapsulates it with TCP/IP. Because of the technology approach that it uses, FCIP's application is limited to interconnecting remote Fibre Channel islands.