In addition to mainstays such as Fibre Channel and parallel versions of ATA and SCSI, Serial ATA is shipping now, and a serial version of SCSI is due next year.
By Alan R. Earls
Like the old Chinese curse, "May you live in interesting times," storage professionals doubtless dream of working in a field that is less "interesting" and more predictable, with fewer complex decisions to make.
Take, for example, disk drives. For years, disk drives were neatly divided between desktop/ portable devices and "enterprise" (server) devices. The former usually relied on the ATA interface, and the latter on SCSI or Fibre Channel. But that began to change a couple years ago and will change a lot more over the next few years.
One of the big drivers in this transition is the fact that parallel interfaces are reaching their performance limits, opening the way for faster, easier-to-implement serial interfaces. For example, Serial ATA (SATA) recently debuted with performance improvements over ATA (which had already begun making inroads into the enterprise market).
Meanwhile, SCSI is also going to get a serial makeover next year (Serial-Attached SCSI, or SAS), and Fibre Channel continues to meet the requirements of high-performance applications.
ATA and Serial ATA
Causing the most stir in the market is SATA, which promises to take desktop drive economics deep into the enterprise space. Jim Porter, president of Disk/Trend Inc., which tracks the disk drive market, says that SATA's market share will increase steadily over the next three to five years. Much of that market share will be in PCs (ATA's traditional domain), but SATA will also make inroads into the server markets that traditionally rely on SCSI.
Porter explains that SCSI and Fibre Channel drives are designed for high reliability and performance, with rotation speeds of 10,000 to 15,000rpm and seek times of only 3 to 6 milliseconds—far faster than ATA/SATA drives. But there are many enterprise applications that simply don't need that level of performance and reliability. Examples include data warehousing, medical image archiving, and any archiving application where high speed and the utmost in reliability are less important than cost and capacity.
In addition to existing applications for which SATA offers a tempting alternative, Porter notes that SATA may be ideal for a number of new applications. "There's a lot of interest in using lower-cost drives for backup-and-restore applications, in effect replacing or complementing tape with a faster alternative," he says.
However, Porter does not see low-cost disk as a wholesale replacement for tape or optical storage but, rather, as insurance against sudden catastrophe or operator error. For applications such as remote mirroring, the high capacity but lower performance of SATA is perfectly adequate.
Porter says: "Remote sites may not need the same performance as the primary site, so why not use ATA drives that cost one-third to one-half the price of SCSI or Fibre Channel drives?"
He also notes that traditional ATA drives have an MTBF of 200,000 to 300,000 hours, while SCSI and Fibre Channel drives typically have MTBF ratings of 1.2 million hours or more.
With ATA and SATA moving toward some of the traditional enterprise markets, some drive manufacturers have begun to raise their MTBF specs on ATA and SATA drives—a move Porter predicts may lead to slight price increases.
John Monroe, an analyst with the Gartner Group, notes that ATA has already proven to be a contender in the lower levels of the enterprise space. In fact, he says, when ATA rotational speeds reached 7,200 rpm, "it killed the low-cost SCSI market."
Dave Reinsel, an analyst with International Data Corp. (IDC), agrees and notes that parallel ATA has completely seized the workstation market and is now chipping away at portions of what once was considered the domain of enterprise-grade SCSI and Fibre Channel drives. That forward march will now be helped by SATA's advanced features, such as hot swapping and command queuing (accumulating read-or-write requests and executing them in the most efficient order).
Reinsel says that traditional mission-critical, transaction-oriented applications will still be the domain of SCSI and Fibre Channel drives, which deliver faster data rates and have higher MTBF numbers. On the other hand, he explains, a lot of existing and emerging applications are just right for SATA. These applications are the same ones described by Porter, so-called nearline storage and archiving applications where the dollar-per-megabyte ratio is crucial.
Still, according to Monroe, while SATA could potentially replace tape libraries, it probably won't because most organizations will still want to have removable media for backup. "It isn't that tape or optical is dead, but that ATA and SATA are being used in effect as a huge cache in front of the archiving media," he says
SATA offers other advantages over its ATA predecessor technology, notes Reinsel. For instance, the lower pin count of its connectors (seven pins for SATA and 40 for ATA) reduces the likelihood of a damaged pin putting a drive out of commission. Then there's the sleek little wire (with a length of up to one meter) that replaces the interstate highway-width flat ribbon cable of ATA, which eases the integration task. SATA also improves on the reliability of ATA drives, particularly in continuous operation mode.
Focusing on the software side of SATA, David Freund, an analyst at Illuminata, notes that SATA can manage a queue of up to 32 commands—a far cry from the 256 commands supported by SCSI and Fibre Channel, but a quantum leap ahead of the one-at-a-time "smarts" of ATA. "At the high end of the market you have RAID controllers that can exploit these features more fully (e.g., changing the order of commands after they've been received)," he says. Freund predicts that SATA's new features will bring "volume economics" into the entry-level and midrange of enterprise storage.
Freund questions to what extent SATA's popularity will alter pricing of enterprise-level drives, if for no other reason than the fact that manufacturers will want to maintain a visible divide between their traditional high-end products and the ATA/SATA drives.
Freund points out that ATA/SATA is no longer based solely on the ideal of commodity products and pricing. He cites Maxtor's SATA drives, which have enterprise-level MTBF ratings and Western Digital's high-revving SATA drives as examples.
SCSI and SAS
In some ways, SATA and SAS are very similar, according to Monroe. For instance, SATA uses the SCSI bus. SAS and SATA even share a physical interface, meaning that systems can be built around either drive type or, potentially, can be converted from one to the other.
But there are many differences, too. For one thing, SAS may have lineage on its side. "In direct-attached storage, there is a rich legacy of SCSI microcode fine-tuned to the peculiarities of various operating systems," says Monroe. "So in certain environments the desire to cling to SCSI is going to be very great," which could confer a market advantage on SAS.
However, Monroe expresses concern about the SAS time line. Though some products may be available in 2004, SAS disk arrays from large OEMs may not be available until 2005. "When you add integration and testing time, it means that SAS systems from major OEMs may not ship until late 2005," by which time SATA may have made crushing inroads into traditional SCSI territory.
Monroe speculates that such a long delay could fatally taint the commitment of top OEMs such as Dell, EMC, Hewlett Packard, and IBM. Their defection would almost certainly be fatal to widespread SAS adoption, he says.
Finally, Monroe argues that the buck will stop on the IT manager's desk, where the budget number crunching happens. Budget constraints won't go away and, increasingly, subsystems that are "good enough" (e.g., SATA) will dominate purchasing requisitions.
But framing the discussion purely in terms of interfaces can be misleading. Discussing the transition from ATA to SATA and from parallel SCSI to SAS may be comforting in its simplicity and linearity but it's only one dimension of the RAID drive story, notes Reinsel. "In the past, drives have been identified and described by their interfaces, but as we move forward there may be a blurring of those distinctions," he says. For instance, enterprise-class drives are expected to run 24x7, with MTBF typically pegged at more than one million hours. They operate at 10,000 to 15,000rpm, and their platters have wider tracks than desktop drives, limiting capacity but reducing access times. Similarly, they are designed for more-demanding tolerances, produce less vibration, and dissipate heat more effectively. Nor is it purely a matter of mechanics. They also have more onboard intelligence, such as command queuing, to enhance I/O efficiency.
Now, with SATA threatening to encroach on the enterprise space, those differentiators may blur. Consider the fact that both SATA and SAS share identical physical interfaces, meaning that in many cases it will now be up to users to select the drives they prefer. And, mindful of the further blurring in the distinction between desktop and enterprise drives that this implies, feature sets are likely to shift.
At least one SATA drive maker—Western Digital—has already delivered drives with 10,000rpm rotation speeds.
Members of the SCSI vendor community, though, are quick to point out that regardless of the interface, a cheap drive is still a cheap drive.
For instance, they point to the tendency of SATA drives to produce more vibration which, when integrated into RAID arrays, can generate harmonics strong enough to disturb read heads and put a RAID box into recovery mode.
But Monroe calls such issues mere speed bumps that SATA drive makers will be able to work around. "When you look at the cost-performance benefits, SATA drives will win as long as manufacturers don't overload the drives with features that would raise prices," he says.
What's the market likely to do with all the interface options? At the high end of the RAID market, Monroe questions whether SATA will make an appearance at all.
"Fibre Channel will continue to dominate the high end" and will continue to gain market share for at least the next two years, he predicts.
Although naturally biased toward their own solutions, vendors are surprisingly unified in their overall views of ATA/SATA, SCSI/SAS, and Fibre Channel positioning.
Sam Sirisena, vice president of worldwide sales and marketing at Promise Technology, says his company has been "telling the industry why ATA drives are good enough for many applications since 1997," and that Promise aims to do the same with SATA.
Commenting on the forthcoming battle between ATA/SATA and SCSI, Sirisena echoes the comments of many pro-ATA vendors: "Good enough is the enemy of the best," he says.
Sirisena also points out that, in addition to the applications already mentioned, Serial ATA will find a home in low-cost, iSCSI-based SAN arrays.
Another factor driving the SATA market, suggests Barbara Murphy, vice president of marketing at 3ware, is network-attached storage (NAS), which has expanded the requirements for lower-cost storage. "SATA is key to [the NAS] market," she says, "and we see SATA playing in the secondary layer between tape and SCSI or Fibre Channel drives."
Joni Clark, Seagate's product marketing manager for SATA, says that Seagate has SATA drives shipping through its reseller channel now and, on the SAS side, will probably be in production in about a year. Clark, who also chairs a SATA industry marketing group, says she expects half of the drives shipping next year to be SATA.
Cautions about sata
However, Linus Wong, director of marketing in Adaptec's Storage Solutions Group, offers cautions about SATA: "People have been seduced by the low cost of ATA." But for RAID, Wong questions the feature sets of ATA and SATA drives. As an example, he cites something as simple as the absence of an LED indicating I/O activity on most ATA/SATA drives. In the desktop environment that might not matter, he notes, "but if you're swapping a bad drive out of a RAID enclosure it's a very helpful feature."
Echoing Monroe's point about the momentum of SCSI, Fujitsu's vice president for advanced product engineering, Mike Chenery, sees SAS products hitting the market in a big way next year. "Because the industry has built so much SCSI infrastructure, we expect SAS to take off relatively quickly," he predicts.
Gartner's Monroe offers a final thought regarding the competing interfaces. Although there are caveats for each, there should be enough growth for all because "the forces that dictate explosive storage growth still outweigh the forces that could inhibit that growth."
So SATA, SAS, and Fibre Channel will, in all likelihood, still be on everyone's radar screens for at least several years to come.
Alan R. Earls is a freelance writer in Franklin, MA.
Storage network, disk and tape interfaces: The basics
By Lee W. Payne
If you're researching interface strategies for your company, you're likely wondering which standards will win, which will die, or if they all will exist side by side. Here's a brief overview of the various interfaces that exist today, or will be available over the next year:
ATA, Serial ATA
Advanced Technology Attachment (ATA), a device-level interface primarily for laptops, desktops, and workstations, is the successor to the original Integrated Drive Electronics (IDE) hard disk interface popularized in the early 1980s. Because ATA drives cost much less than Fibre Channel or SCSI drives of equal capacity, over the next few years disk arrays will increasingly incorporate ATA interfaces to deliver more cost-effective storage to price-sensitive markets.
Serial interfaces, such as Serial ATA, are faster than parallel interfaces and simplify RAID controller architectures by reducing the board-level routing nightmares associated with parallel buses. Serial buses also are easier and less expensive in switching environments.
The Small Computer System Interface (SCSI), the dominant protocol for controlling block-level access to data, has become the mainstay of enterprise-class storage systems. Fibre Channel, for example, uses SCSI as an upper layer (FC-4) protocol. Because all of the block-level storage protocols for Gigabit Ethernet—iSCSI, iFCP, and FCIP—use SCSI Command Descriptor Blocks (CDBs), SCSI CDBs will remain the key software protocol for enterprise-class storage for the foreseeable future.
Parallel SCSI is an ideal interface for direct-attached external storage. The next logical progression in SCSI's more than 20-year run is Serial SCSI, also known as Serial-Attached SCSI (SAS). Serial SCSI will be a low-cost, high-speed serial interface, with products expected next year. What remains to be seen is whether Serial SCSI can become a cost-effective alternative to Parallel SCSI. Serial SCSI will also compete with ATA and Serial ATA at the low-end (desktops, workstations, and low-end servers), and with Fibre Channel and Serial ATA in higher-performance disk arrays.
For more information on SCSI, visit the SCSI Trade Association at www.scsita.org.
As an interface for tape libraries, Fibre Channel will continue to replace SCSI as companies begin to share their backup resources through storage networking technologies such as Fibre Channel storage area networks (SANs). However, because Fibre Channel disk drives command a 4x to 5x price premium over lower-cost disk interfaces, Serial ATA and Serial SCSI may erode Fibre Channel's market share in the back-end disk array space. But Fibre Channel will continue to be a dominant front-end interface for the foreseeable future.
An InfiniBand fabric uses a switched fabric topology similar to Fibre Channel and has three key components—Host Channel Adapter (HCA), switch, and Target Channel Adapter (TCA).
An HCA connects a host system to the InfiniBand fabric. An InfiniBand switch connects HCAs to HCAs or TCAs, and a TCA connects non-initiating devices to the InfiniBand fabric.
InfiniBand has significant potential as an interface for blade server clustering because of its low latency, simplified serial interface, and well-defined management capabilities. It also has potential in the high-performance computing (e.g., supercomputer) market.
Gigabit Ethernet (GigE) will have a profound effect on storage going forward because it provides the hardware and Ethernet-TCP/IP foundation for several higher-level protocols associated with data movement and backup, such as NDMP, iSCSI, FCIP, and iFCP. All of these storage-over-IP standards share a common hardware platform in addition to the underlying TCP/IP protocols.
One key advantage that GigE has over previous generations of Ethernet (10Base-T and 100Base-T) is that it permits only switched traffic flow. GigE does not permit hubs and eliminates delays caused by collisions, yielding a more predictable throughput.
The Network Data Management Protocol (NDMP) is a higher-level protocol for Ethernet that is often used for backup of high-end network-attached storage (NAS) filers. A tape library that supports NDMP version 4 enables data to be transferred directly from an NDMP-compliant NAS device to a library under the control of the backup server running NDMP client software.
Internet SCSI (iSCSI) is an emerging protocol for transferring block-level storage I/O over standard Ethernet networks, resulting in what is often referred to as IP SANs. iSCSI will enable cost-effective sharing of tape libraries over GigE LANs to help resolve backup problems. Tape drives and libraries appear as locally attached subsystems to distributed servers—at a lower cost than with Fibre Channel. (For more information, see "iSCSI gains a toehold in SAN market," InfoStor Special Report, April 2003, p. 18.)
Fibre Channel over IP (FCIP) encapsulates Fibre Channel in IP packets and provides connectivity between "islands" of Fibre Channel SANs across MANs or WANs using TCP/IP. FCIP relies on standard GigE hardware for switching and routing. FCIP devices are essentially protocol bridges between Fibre Channel and GigE, providing the hardware and firmware to encapsulate Fibre Channel frames into TCP/IP packets, and vice versa.
The Internet Fibre Channel Protocol (iFCP) provides SAN-like fabric functionality based on Ethernet switching. It is a TCP/IP-based protocol that allows a SAN to be created with Fibre Channel end devices and GigE switching and routing in between. iFCP devices, or "gateways," provide both switching and bridging functionality.
10 Gigabit Ethernet
10 Gigabit Ethernet (10GigE) will provide a data-center backbone and will become the point of hardware convergence between block-level SANs and file-level NAS. It will support block-level storage protocols such as iSCSI, FCIP, iFCP, and NDMP, as well as file-level protocols such as NFS and CFS. 10GigE will also be used as a switch-to-switch pipe. Today, however, the cost for a single 10GigE port is more than $10,000.
Lee W. Payne is a senior system technology strategist at Overland Storage (www.overlandstorage.com) in San Diego, CA.
Serial ATA II Native Command Queuing
By Stephen P. Leo and Amber Huffman
As Serial ATA (SATA) gains momentum, one of the most anticipated features is the SATA II Native Command Queuing. This feature adds data-handling intelligence that is expected to deliver the performance needed for the next generation of entry-level servers, networked storage, and high-end PCs.
How does this new feature work, and how can it benefit you? First, it is important to understand command queuing basics and then to see how three key elements of Native Command Queuing build on that foundation to deliver even higher performance.
In general, command queuing enables a hard drive to accept multiple commands from the host controller and then rearrange the order of those commands to maximize throughput. Most of a drive's command service time is taken up by seek and rotational delay while the drive head lands on the appropriate data to transfer. Command queuing can reduce this time by overlapping drive seek and rotational delay with the command issue and completion overhead.
For example, at the same time a new command is issued to the drive, the drive can seek to locate the data for a different command. Also, the drive can optimize the selection of the next command to service, based on the location of the command's data relative to the head's current location.
In the communication between host controllers and disk drives, tags are assigned to each queued command. To transfer data, the drive communicates command tags to the host. The host then sets up a Direct Memory Access (DMA) operation that points to the appropriate host memory region for the command based on the tag value.
In the completion phase, the disk drive sends the data to the host controller, which writes the data to the host memory region. The drive also sends status information to update the host controller on the status of pending commands or errors.
Benefits of Native Command Queuing
The SATA II protocol extensions streamline the command queuing and communication process. Increased performance and efficiency primarily result from three new capabilities: race-free status return, interrupt aggregation, and First Party DMA.
Race-free status return allows status to be communicated about any command at any time. There is no "handshake" required with the host for this status return to take place. The drive may issue command completions for multiple commands back-to-back or even at the same time.
In theory, a drive can interrupt the host each time it is ready to complete a command. The more interrupts, the bigger the host processing burden. However, with SATA II Native Command Queuing the number of interrupts per command is usually less than one. If the drive completes multiple commands in a short time span—a frequent occurrence with a highly queued workload—the individual interrupts may be aggregated. In that case, the host controller only has to process one interrupt for multiple commands.
SATA II Native Command Queuing also has a First Party DMA (FPDMA) mechanism that lets the drive set up the DMA operation for a data transfer without host software intervention. The drive selects the DMA context by sending a DMA Setup Frame Information Structure (FIS) to the host controller. The FIS specifies the tag of the command for which the DMA is being set up. Based on the tag value, the host controller will load the table pointer for that command into the DMA engine, and the transfer can proceed without any further setup.
Serial ATA II Native Command Queuing optimizes the command queuing mechanism, delivering enhanced performance. It achieves this performance through a streamlined protocol, using hardware to automate DMA setup and minimizing the number of interrupts required for data transfers.
Drives and host controllers with SATA II Native Command Queuing will be fully interoperable with current drives and controllers. New Native Command Queuing devices can be phased into the existing infrastructure over time. As devices begin to support SATA II Native Command Queuing later this year, organizations can take advantage of this enhanced performance they will need for their next generation of entry-level server, networked storage, and high-end PC products.
Stephen P. Leo is a technical marketing engineer in Intel's Storage Components Division, and Amber Huffman is a staff software architect in Intel's I/O Architecture and Performance Division.