Making Connections

Hard disks could have infinite capacity and instant read and write speeds, and still wouldn't be of any use without some way to communicate with the rest of the computer system, which is the job of the interface. But simply providing a path for communications isn't enough. The interface also has to be fast enough to keep up with the drive. Otherwise it would be a bottleneck that prevented increases in hard disk performance from making any difference. This need for the interface to stay out of the way is why interface speed matters, and why interfaces keep getting faster as hard disk performance improves.

Up until a few years ago there were only two interfaces that mattered for most purposes: ATA (Advanced Technology Attachment) and SCSI (Small Computer System Interface). Today, there are three more: USB (Universal Serial Bus), IEEE 1394, and Fibre Channel, with SATA (Serial ATA) on the near horizon.

At the enterprise level--where NAS (Network Attached Storage) and SANs (Storage Area Networks) are increasingly more important than DAS (direct attached storage; Read: a hard disk in a computer)--iSCSI (Internet SCSI) and InfiniBand are looming as key alternatives to fibre channel as fabrics for enterprise storage needs. (A fabric consists of the hubs, switches, and software that tie the pieces together.) Here is a quick look at future of interfaces and fabrics, with an emphasis on the interfaces, and on the desktop.

ATA--also known as IDE (Integrated Drive Electronics) or EIDE (Enhanced IDE)--has always been a parallel interface, moving multiple bits of data in parallel over a flat, ribbon cable. The current standard for ATA is ATA/100, which has a 100 MB/sec maximum throughput.

Lately, the industry has started to call the traditional ATA interface "parallel ATA"--and some abbreviate it to PATA--to distinguish it from the upcoming Serial ATA (SATA). (The final version of the SATA 1.0 specification was released on August 29, 2001 and is available at www.serialata.org)

Aside from the increase to a top speed of 150 MB/second in the first generation specification, SATA offers several advantages over parallel ATA. In particular, by replacing the broad, flat ribbon cable with a thin serial cable, SATA allows better airflow for cooling--an important consideration in today's systems, with their (metaphorically and literally) hot components, and an even more important consideration for future systems that may want to avoid cooling fans.

Another benefit for SATA is simpler installation, with one drive per cable and no more unwieldy ribbon cables. There's also the potential for hot-plugability--an issue that matters more on the enterprise level than the individual desktop. A strong plus for SATA in any context, finally, is that it's compatible with operating systems and other software written for parallel ATA. As Intel's Jim Pappas puts it, it keeps the ATA interface, and just changes the physical layer, which makes it transparent to software that was designed for ATA. And that means you can use it even with, for example, an older version of MS-DOS.

Until recently, SATA was universally seen as the next step after ATA/100. But since Intel announced that it would not provide a chipset with SATA support until sometime in 2003 (in its ICH5 chipset component), some companies--including drive vendor Maxtor and chip vendor Via Technologies--are opting for a slight detour, by way of ATA/133, a 133 MB/second version of parallel ATA.

The debate is between those who feel that ATA/100 will turn into a performance bottleneck before SATA is available, and those who don't. Companies in the first group--most notably Maxtor--are already moving forward to ATA/133. (At this writing, Maxtor has announced an ATA/133 drive model.) The graph below, which was provided by Maxtor, shows the interface rate needed (the green line) for any given drive data rate (the red line), and maps both over time. Note that the required interface rate goes above 100 MB/second, indicated with the line labeled Ultra ATA/100, in early 2002. And also note that the required interface rate is higher than the drive data rate due to host and protocol overhead.

Interface transfer rate needs to be greater than the HDD disk data rate to avoid becoming a bottleneck Interface transfer rate faster of ATA/133 essential for 2002 With host overhead, bottlenecks can start at 60% of the interface rate.

Those in the second group--including Intel and Seagate--feel that SATA will be available soon enough, and that ATA/133 isn't worth the investment of time, money, and energy it would take. SATA should be available in PCI cards in early 2002, and on motherboards in the second half of 2002, but with data still going though the PCI bus and limited to the PCI bus throughput speed (which tops out around 70-90MB/sec in practice instead of the theoretical max of 133MB/sec, much as bottlenecks in the ATA interface start at about 60 percent of the interface rate). Seagate says it expects manufacturers to offer SATA drives by mid-2002, but doesn't expect them to be pervasive until 2003. The chart below shows the expected SATA timeline.

Ultimately, this split in the industry is a decidedly minor issue. Even those who are planning to move to ATA/133 are also planning to move to SATA when it's available. Far from being a fork in the road, the two different paths are more like driving down route 100 and seeing a sign for route 100A, as a minor detour that will reconnect with 100 further down the road. In short, the roadmap for SATA represents the future for parallel ATA as well, with a new generation every three years to double the top speed or the interface, as shown in the chart.

You may wonder about the new 3GIO (Third-Generation I/O initiative) and how it fits into the picture. As far as we can tell, based on what we've heard from 3GIO proponents, although 3GIO will be able to connect a variety of products over high-speed serial interconnects in the future, SATA has reduced latency and reduced protocol overhead. And SATA is not general purpose, but rather specific to storage and is 100% ATA software compatible, unlike the 3GIO interface. For more insights into 3GIO and other future interfaces, read Nick Stam's Fall IDF 2001 Report.

As SATA grows faster and becomes more pervasive, it's likely to challenge SCSI on the power user desktop, and even in servers, but SCSI still has plenty of life left. Or, as Seagate puts it, predictions of its demise are premature. Seagate says that Ultra320 SCSI (with a top speed of 320 MB/sec) is much more robust than expected, Ultra640 SCSI "looks feasible," and Ultra1280 SCSI may lie down the road as well. In any case, the future for SCSI over the next few years is straightforward: More of the same, growing faster with time, and staying ahead of SATA. The chart below shows the SCSI timeline, based on likely dates for early evaluation products according to Seagate.

External Drives USB 2.0 and IEEE 1394 (which is also known as FireWire and iLink) are more directly competitive with each other (at least as far as hard disks are concerned). Both interfaces are good choices for external hard disks, with current top speeds of 60 MB/sec for USB 2.0 compared to 50 MB/sec for IEEE 1394. At this writing, there are no clear plans for the next step beyond USB 2.0, but IEEE 1394b can handle up to 3.2 Gbits/sec, or 400 MB/sec, and IEEE 1394 drive interfaces are expected to operate at 200 MB/sec by 2005. The chart below compares USB 2.0 and IEEE 1394 to both parallel and Serial ATA.

Interface Parallel ATA Serial ATA Universal Serial Bus(USB 2.0) IEEE 1394(FireWire or i.LINK) Target Market Internal storage devices Internal storage devices External devices, (keyboard, mouse, printer, scanner, storage) External storage devices Cost Baseline Equals parallel Greater than ATA Greater than ATA Speed in Year 2000 (MB/s) 100 N/A 60 50 Projected Speed in Year 2002 (MB/s) 133 150 60 50 Projected Speed in Year 2005 (MB/s) Obsolete? 300 Unclear 200 Hot Plug No Yes Yes Yes

At the enterprise level, SCSI has traditionally been the drive interface of choice for direct-attached storage and fibre channel has been the interface of choice for the SAN network infrastructure. In large SAN installations, fibre channel drives are also common. NAS tends to use SCSI drives, and, to a lesser extent, ATA. Neither SCSI nor fibre channel drives will disappear any time soon, but Intel expects SATA to challenge both. Mike Wall at Intel expects SATA to first replace parallel ATA, then replace at least some SCSI and even some Fibre Channel applications at some point.

Note that iSCSI and InfiniBand do not define a storage device interface as fibre channel does (at least not yet). They are both used in switches and hubs, connecting to drives that use any arbitrary interface. They are also implemented in servers for direct connection to SAN storage subsystems and to other servers in the case of InfiniBand, and to remote storage resources over the IP network in the case of iSCSI. They represent a challenge to the fibre channel networking fabric rather than fibre channel drives. Until now, the fibre channel fabric has been the only game in town for SAN applications. And, indeed, iSCSI and InfiniBand promise to chip away at various levels of the SAN market. However, most fans of these two options see iSCSI less as a direct competitor for fibre channel than as a complementary choice. iSCSI provide more affordable storage networking, and great flexibility in applications that can't justify the cost of fibre channel and don't require the same performance.

For more details on SANs, iSCSI, and Fibre Channel, see our story Introduction to SANs: Technology, Benefits, and Applications. We plan to have in-depth stories on InfiniBand and iSCSI coming soon.

The three graphics below provided by Seagate show the expected timelines for fibre channel, iSCSI and InfiniBand development, along with some of the key points relating to each. As Seagate's graphic shows, InfiniBand was originally designed to replace PCI, but we all know that 3GIO now fits that bill on desktop systems, and InfiniBand will be used to interconnect servers with each other and with external storage. Chipsets for the upcoming McKinley IA-64 chip will have InfiniBand support included.

In the context of all drives shipped world wide for all markets, Dave Reinsel of IDC projects that as Serial ATA grows in market share, the percent of SCSI drives will shrink noticeably even by 2005, as the chart below shows.

Copyright © 2004 Ziff Davis Media Inc. All Rights Reserved. Originally appearing in ExtremeTech.