Serial ATA is somewhere between 'desktop' ATA and 'enterprise' SCSI and Fibre Channel disk drives, but will drive makers continue to classify (and price) disk drives based on interface technology?
By David Freund
Serial ATA (SATA) is steadily gaining ground—and not just in desktop PCs as a replacement for the tried-and-true parallel version of the venerable ATA disk bus and its IDE predecessors. SATA has now begun to infiltrate enterprise data centers, showing up in low-end and even midrange storage arrays. With storage costs becoming an increasing percentage of IT spending, the trend is hardly surprising since SATA drives are priced significantly lower than SCSI and Fibre Channel drives. Pricing varies widely, but end users can expect to pay 30% to 50% less—on a per-megabyte basis—for Serial ATA disk arrays versus Fibre Channel disk arrays.
Despite its compatibility with the heretofore very "low-end" ATA protocol, SATA is a different animal. It greatly improves not only on ATA's "speeds and feeds" and physical wiring flexibility, but also incorporates many of the capabilities and optimizations previously available only in more enterprise-class SCSI and Fibre Channel alternatives.
At the same time, Serial-Attached SCSI (SAS) will rise to replace its older, parallel predecessor—SCSI—which currently dominates data-center drive sales. Drive manufacturers are positioning SAS to take SCSI's place in the market.
The distinction between SATA and SAS capabilities, however, is considerably less than the differences between the previous generations of SCSI and ATA. While SAS has some features that SATA still lacks, economic interests are the principal reason that manufacturers continue to segment disk drives—and prices—along SATA/SAS boundaries. But economic interests could well be manufacturers' undoing. It takes only one dissenter among the major drive manufacturers to show the storage-buying world just how artificial that distinction has become.
The chicken and the egg
ATA drives are everywhere. They pervade the PC market, providing excellent value and enormous capacity, if less than inspired performance. The catalyst for Serial ATA to break out of its PC niche and into the data center is the significant leap in capability that SATA has made over parallel ATA.
ATA drives were effectively chained to the desktop for a simple reason: The parallel ATA bus is too limited to handle any more workload than one desktop user can generate. Because of the ATA bus' nature, performance and reliability requirements for ATA drives have always been significantly lower than the requirements for SCSI and Fibre Channel drives. The costs associated with designing, building, and testing drives that meet those different sets of requirements made the disparity in prices among the drives intuitively understandable to customers. More important, the distinction by customers of the value between the two drive types was both real and considerable.
The SATA interconnect and protocol eliminates several of ATA's performance restrictions. Electrically, moving to a serial interconnect eliminates signaling noise problems. It also enables a much smaller and more flexible cable, making physical routing within a system enclosure much easier. (These advantages, by the way, are shared by SAS because both interfaces share the same physical layer.) But SATA also eliminates ATA's master/slave bus sharing, which allowed only one drive to communicate at a time. SATA uses a point-to-point (vs. bus) connection, providing the ability to guarantee bandwidth to each drive. SATA also adds support for hot-swapping drives—which was not possible with parallel ATA drives. Reliability is improved by adding cyclical redundancy checking (CRC) for both data and control information (parallel ATA only checked the data path). And bandwidth is boosted from 133MBps to 150MBps.
It is I/O latency, however, where SATA gains the most over its predecessor. Unlike desktops, data-center environments tend to have heavy loads of relatively small—but extremely random—I/Os. In those environments, it's not bandwidth, but latency, that matters most in delivering application performance. A major contributor to SATA's boost in I/Os per second (IOPS) performance is command queuing. Multiple commands are issued to a drive, which then completes the commands in the most efficient order. Though not well-known, even Ultra ATA supports tagged command queuing (referred to as ATA TCQ). However, it has been largely avoided by drive and system manufacturers. And for good reason: The implementation had inefficiencies that could actually hurt performance for lightly loaded single-user desktops. SATA replaces ATA TCQ with native command queuing, efficient command-completion handshaking, and first-party DMA (direct memory access) to reduce latency issues. This opens up many server opportunities for SATA that previously were out of reach.
What about SAS?
But SAS delivers the physical-layer benefits of SATA, and more, raising the issue of relative positioning of the two serial interfaces.
SATA can manage a queue of up to 32 commands. That's a significant advance, but it remains a far cry from the 256 commands supported by SCSI (including SAS) and Fibre Channel. SATA drives have only one I/O port; SAS drives will be dual-ported, providing an extra path to each drive in the event of a controller failure. (The two ports can also be combined to create a virtual "wide" port. When connected to two ports of a SAS controller that supports port aggregation, the effective I/O bandwidth can be doubled.) SATA drive ports are half-duplex; SAS ports are full-duplex, meaning each port can receive and transmit information simultaneously. A SAS host port can (eventually) support up to 200 devices, whereas SATA supports just 16. This adds up to three advantages for SAS: better per-drive IOPS performance under very heavy loads, additional qualification requirements specific to those loads (as well as dual-ported configurations), and better per-port scalability.
Although SAS has some technical advantages, the key question is: How much do these advantages matter for the vast majority of the drives spinning in today's data centers, now or in the future? Networked storage is increasing as direct-attached storage (DAS) use declines, and networked-storage controllers have a considerable amount of intelligence built into them. They already present "virtual volumes" to connected servers, spreading and optimizing I/O loads over multiple drives. Only extreme top-end array designs really need the advantages of such deep per-drive command queues, or huge numbers of connected drives per port. The majority of stored data in commercial environments is not accessed with anywhere near that intensity. As for servers, in an increasingly networked-storage environment, server port counts will tend to gravitate toward a fairly low number—one that is completely unrelated to what interfaces are on individual disk drives.
The fact is, the most important factor affecting disk-drive performance has little to do with the interface used, but, rather, with the electro-mechanical components used. The historical model of reserving the latest, fastest, and most-reliable designs for SCSI and Fibre Channel (thus leaving older, less-reliable mechanicals for the ATA space) made sense in the parallel-interface generation. But for most workloads, the interface is largely diminished as a distinguishing factor. And it doesn't differ much in production costs, either. From a technical standpoint, there's nothing standing in the way of drive manufacturers combining their best electro-mechanical components with SATA interfaces.
This is not to say that all drives should become equal. The requirements for disk drives in a desktop PC differ significantly from those in a data center. First, desktop PCs are typically used by a single user. Although each user may have many windows open, it's rare for more than one or two applications to read and/or write appreciable amounts of data on the PC's hard drive simultaneously. Second, most PCs are actively used only eight hours per day, five days per week, and are either idle or turned off the rest of the time. Third, desktops usually contain only one or two drives and are usually located in a relatively quiet office environment.
Data centers, on the other hand, are a much harsher environment. Drives are accessed by multiple users seeking information from random parts of the drive simultaneously, making the magnetic-head actuators work much harder. The better a drive is at optimizing the order in which it executes multiple commands, the better the performance. That much activity also generates more heat and more wear and tear on drive components. On top of this, enterprise drives are usually stuffed into multiple rack-mount shelves, concentrating the heat generated by those drives in a confined space. Powerful fans are typically part of the enclosures, solving some of the heat problem, but adding to vibration and noise. Since these parts usually get bolted together to the same metal frame, the vibration from the various components is transmitted to others.
If not mitigated, vibration can pose serious danger to data stored on the drive, since there's not a lot of room for mechanical error in today's disks. Unlike other forms of magnetic media such as tape, where read/write heads physically touch the surface of the recording medium, the head of a disk drive "flies" only one-half micron above the magnetic-disk platter. If a head touches the spinning platter (which is spinning many thousands of times per minute), it will carve a deep groove across the surface, destroying the drive's usability—and data.
As a result, drives meant to be used in enterprise data centers require more exacting and resilient designs and more rigorous qualification. Both add to the cost—and value—of an enterprise drive. Manufacturers also tend to adapt their warranties to suit the drive type. Most have "mean time between failure" (MTBF) ratings in the million-hour range, but warranties can be restricted based on the percentage of the time a disk is actively reading or writing data, referred to as the "duty cycle." Desktop drives are typically warranted to operate in "light duty-cycle" environments, ranging from 10% to 20%. Enterprise drives, in contrast, are rated for 100% duty-cycle environments.
In my opinion . . .
The link between the drive interface—SATA or SAS—and whether a drive is "enterprise class" is no longer valid. Except for conditions of extreme I/O load (seen only in a small minority of applications in commercial data centers), the gating factors for a drive's performance and reliability are in its electromechanical components. It makes sense for drive manufacturers to offer different classes of drives for different environments, such as single-user desktops, archival arrays, and high-performance data-center arrays. But for the vast majority of applications, it no longer makes sense to base a drive's type on its interface.
So why do drive manufacturers persist in doing so anyway? In part, it's because that's the path of least resistance. However, the more likely reason has more to with protecting profit margins. Desktop drives, once synonymous with "ATA," represent three-fourths of total disk-drive shipments and have been under tremendous price pressure, forcing razor-thin margins for manufacturers. Enterprise-class drives are more expensive to design, build, and qualify, but manufacturers of those devices make much more profit on them—a situation they desperately want to protect.
Over the past 15 years, the disk-drive industry consolidated significantly, reducing the number of major players from more than 20 to just seven (Fujitsu, Hitachi, Maxtor, Samsung, Seagate, Toshiba, and Western Digital) with the top four (Seagate, Maxtor, Hitachi, and Western Digital) commanding more than 80% of the market. Maxtor's acquisition of Quantum's disk business in 2001 removed one of the last aggressive price cutters from the market, making for what some have called a "more rational pricing environment." Combine that with technology improvements that have lowered the component count on each drive, and margins have been gradually improving.
There's an ever-increasing thirst for storage capacity, and the number of units shipped (even with the capacity of each increasing) is also expected to rise steadily. So things should be good for drive manufacturers, right? Not quite. Continually decreasing disk prices would turn this into a zero-sum gain for the industry overall. If drive makers all agree on positioning their drives according to existing practices, then the industry as an aggregate can increase its revenues. While obviously constrained from formalizing such agreements, this is not too far from the model of OPEC nations agreeing on oil quotas and setting prices.
The song sheet goes something like this: SATA is positioned (or classed) as low performance and low reliability, suitable for infrequently accessed data or backups, but not much else in a data center; SCSI and SAS are positioned as the high-performance, high-reliability choice; Fibre Channel comes in as an odd sort of "top-shelf;" it's not really a better performer than SCSI, but it commands top prices.
Systems vendors (both servers and networked storage) are likely to embrace both SAS and SATA by incorporating SAS on their motherboards, since SAS is backward-compatible with SATA. It's a safe move; they'll be able to offer both technologies, with lower costs (and fewer SKUs, which resellers love), and let the volumes fall where they may.
But, like OPEC, it takes only one member to spoil the party for everyone else by cutting prices to increase market share. And in a zero-sum market, the only way an individual vendor can grow its revenues is to grow its market share. Seagate, for example, would be loathe to commit such an act, with about half of its revenues coming from sales of SCSI and Fibre Channel drives. But a Maxtor or Western Digital has much less to lose—and more to gain—by using SATA to grab more market share.
SATA makes its move
Western Digital has begun this very attack with pitches like "all the speed and reliability of SCSI without the wicked price tag" in the sales literature for its WD Raptor drive. The company has taken an electro-mechanical design meant for long-term vibration-prone, rack-mounted use (sounds like "enterprise," doesn't it?) and put a SATA interface on it. Other vendors are quick to point out how SATA is really a "desktop technology," and end users have tended to buy that story. So far.
But there are other vendors higher up the storage food chain that are getting impressive performance from these "lowly" SATA drives. Panasas, for example, uses SATA drives in a clustered-network-attached storage (NAS) configuration (with Celeron processors) and has been able to set records in NAS I/O performance. Even though a few competitors can get similar results, they can't match the low dollars-per-gigabyte or dollars-per-I/O levels Panasas achieves. As has occurred in the server market, the threat—or promise, depending on your point of view—of volume economics is looming over data-center storage.
If the group of seven drive manufacturers succeeds in deflecting such potent market forces, their practices would certainly deserve closer study by business schools...and perhaps government officials. If, on the other hand, the drive manufacturers abandon their artificial margin protection and compete directly on attributes that really matter to end users—performance, reliability, capacity, and price—they'd make it easier for users to discern which products offer the best value. And let the interface volumes fall where they may.
David Freund leads the Information Architectures practice at Illuminata Inc. (www.illuminata.com) in Nashua, NH.
For more information…
about Serial ATA, visit www.serialata.org.
about Serial-Attached SCSI, visit the SCSI Trade Association's Website, www.scsita.org.
Related articles in InfoStor:
"Tech preview: Serial-Attached SCSI + PCI Express," p. 26
"Serial ATA activity heats up," January 2004, p. 1
For a review of recent SATA product announcements, see the New Products section, p. 45.