2.2 Storage Attachment
Now that we've listed a large number of devices that go into building a storage network, let's consider how these devices will be connected and accessed. Perhaps one of the most interesting aspects of storage networking is the proliferation of protocols and attachment techniques. For example, a higher-level protocol such as SCSI can be routed over a typical IP network.
In general, the three storage attachment types mentioned here can be considered "independent" of any protocol. The first attachment type, direct attach storage (DAS), is the traditional storage on the host's bus implementation. This attachment type can be seen in any home computer and in most file servers available in small to medium businesses. The next typesnetwork attach storage (NAS) and storage area networks (SANs)decouple the storage from a single host. There are two major differences in the implementations of SANs compared with NAS devices. One is that NAS storage tends to use existing networking infrastructure, whereas SANs use networks specialized for storage data. The other is that SANs specialize in moving block data, whereas NAS is optimized for file request and fulfillment.
2.2.1 Direct Attach Storage
Direct attach storage can be considered the traditional mechanism for attaching storage to a network. If you are on a single-user system or a traditional host, you likely have drives attached via an IDE bus or a parallel SCSI bus. This makes the storage subservient to your host. This means that if your host is unavailable, the storage attached to it is more than likely unavailable.
Although bus attached storage is adequate for some businesses, it begins to fail for multiple reasons when the storage is scaled to enterprise-class systems. Host systems can easily become bottlenecks in a storage network. By decoupling storage from the host, you alleviate bottlenecks by giving multiple access points to data. Furthermore, direct attached storage implies tying your storage to a single system. If that system fails, your files are unreachable. Decoupling the storage from the hardware allows many more failover and replication options.
Backup and archiving also becomes a problem. Data must flow through the host from the source to the target location, most likely a tape device. This configuration creates a bottleneck on the bus and puts a heavy load on the server. If the tape device or backup program is physically separate from the machine on which the storage resides, severe network congestion can occur.
Partitioning storage from the host can alleviate these issues by optimizing attachments between data suppliers and data consumers if they are all attached to the storage network. Bus attached storage also has a built-in distance limitation. The longest parallel SCSI implementations are approximately 25 meters with special cabling. By removing storage from the cabinet and bus (in which you are no longer direct attached), you can relieve these distance limitations. For example, the fibre channel protocol used to implement storage area networks supports cabling of up to 10 kilometers at 1 gigabit/second transmission rates. Unlike the bus attached storage, with its built-in limitations, the distance and bandwidth supported by the fibre channel protocol enable far superior data center designs that can easily be scaled up.
2.2.2 Network Attach Storage
Network attach storage devices come in a wide variety of flavors that can be used in applications ranging from personal use all the way up to enterprise-size businesses. Prices range from a few hundred dollars to hundreds of thousands of dollars. A NAS device is basically an optimized file server that can be reached by a traditional remote drive assignment (such as NFS or CIFS).
Typically, NAS devices are purchased for the purpose of sharing files among multiple clients and even multiple operating systems. NAS devices do not address networking issues or backup issues because the device is attached directly to a production network.
Onboard the NAS device is an operating system specialized for file handling that manages network connections and the storage that is onboard the NAS device. Most NAS devices come with more than one internal drive. These drives can be configured in several ways, including different RAID levels, so as to optimize storage for the particular QoS attributes that are most desired. How the RAID levels are implemented differs depending on the price of the device. For example, a low-end NAS device likely uses software-based RAID built into the onboard operating system. A higher-end RAID device probably uses onboard RAID controllers.
Low-end NAS devices fall into the category of shared disks. High-end NAS devices include tape libraries and sophisticated HSM software to make offline data retrieval seamless and unnoticeable by an end user. If a low-end NAS device is purchased, the system administrator will probably need to find a way to achieve data backup, possibly through Network Data Management Protocol (NDMP) implementation that the device has onboard (NDMP is discussed later in this chapter). Unfortunately, not all low-end NAS devices are NDMP-capable.
2.2.3 Storage Area Networks
A storage area network is a specialized network used primarily for storage traffic. Whereas NAS devices are self-contained file-serving units installed on an IP network, SAN devices are isolated to their own network and require a host to surface the data to clients on the production network. SANs are made up of network infrastructure components that function similarly to typical IP network components, and end points such as RAID devices and tape libraries.
SANs are closely associated with the fibre channel protocol, although other SAN implementations are becoming available that use gigabit ethernet. Fibre channel is a SCSI I/O protocol that typically runs over two types of cabling schemescopper or fiberat very fast transmission rates. Fibre channel supports transmission distances of up to 10 kilometers and effective data rates that far exceed those of SCSI and IP at the time of this writing.
Recall that a NAS device is typically accessed at the file level. SANs are typically made up of block devices. Hosts surface files back to the client that made the file request; the host system receives a request for a file, and the file system on the host passes requests for blocks of data to individual devices and addresses on the SAN. The host reassembles the block data and returns the requested file to the client.
Multiple hosts can access block devices on a SAN, but the sharing is at a block level; in contrast, NAS devices share data at the file level. Also, substantially different architectures are required, but great advantages can be had by using a SAN rather than one or more NAS devices, especially when a data center grows past the terabyte range.
Although we've attempted to draw a distinction between NAS and SAN, that distinction seems to be fading. The storage industry is on Internet time, and the lines between protocols and hardware types are blurring all the time. NAS devices will be produced that will allow block access, and SANs are being produced that look more like NAS devices.
Fibre channel has three topology options, only two of which are actually storage network topologies:
Point-to-point. This is basic host-to-device configuration. This attachment option actually fits better in direct attach storage, although it is usually placed with SAN because of its fibre channel implementation.
Arbitrated loop. A shared loop that all devices participate in, this is a low-cost solution for creating storage area networks. Arbitrated loops are typically built with hubs and block devices. The downside of using arbitrated loops is that they tend to be difficult to debug, mainly because the loop must be broken to insert special equipment for debugging. The very nature of breaking the loop and inserting a device that introduces a different timing characteristic may make the problem undetectable. Another problem is that if a break occurs in the transmission loop, all is lost until the break can be found and replaced. You can alleviate this risk by running redundant loops. If this is a concern, a better approach is a switched fabric. Finally, arbitrated loops share bandwidth between all devices participating in the loop. Switched fabrics do not have this restriction.
Switched fabric. A switched fabric is probably the most dynamic and impressive of the storage networks. Switched fabrics tend to be expensive because they require networking infrastructure, but their manageability and debug capabilities can make them worthwhile. Also, as the components required for a switched fabric come down in price, so will the overall cost of this topology. Network switches can detect devices that are disruptive and switch them out of the fabric until they can receive the attention they need. Rather than a loop topology, the fabric is based on a star topology. Loss of a connection to a single device, or loss of a device, does not have as severe an impact as does the loss of a device in an arbitrated loop.
By partitioning storage infrastructure into its own network, you enable an amazing amount of optimization. Furthermore, storage devices can maintain their own communications paths within the storage network, thus keeping high-bandwidth, local usage, such as backup and archiving of data, off of the production network (the network carrying client-to-client communications such as TCP/IP traffic). Although fibre channel is usually found in the same sentence as SAN (or at least implied by the term SAN), there is no reason that a storage area network could not be implemented with a gigabit ethernet. In fact, in the coming years we will likely see storage networks using much of the same infrastructure as high-speed production networks. Using a single set of management tools to manage both storage and production networks may become a reality. FMA is abstract enough to handle any type of network topology, hardware, and software.
In general, SAN solutions, as well as generic storage networks, require a great deal of planning before implementation. On the other hand, the versatility and optimizations that can be achieved by separating data into a separate storage area network can easily offset costs and create a more robust, dynamic, and flexible environment.