Years 2004 and 2005 are widely expected to exhibit rapid acceptance of the Serial ATA (SATA) storage interconnect standard. Hard drive manufacturers are busy developing SATA products with advanced “SCSI-like” features, such as Native Command Queuing (NCQ), in-band power management, high spindle speeds and, of course, massive storage capacities. While 3.5” form factor parallel ATA drives have predominantly found homes in the commodity desktop market, and their Serial ATA successors are expected to dominate this market, we believe that the speed and simplicity of the SATA interface also portends lofty ambitions for the unassuming little 2.5” drives up to now found only in mobile platforms. The inclusion of high performance elements in SATA drives, of limited usefulness in commodity desktops, presents an inevitable challenge to SCSI and fibre channel (FC) hard drives in enterprise storage systems.

Storage Capacity for the Desktop

Let’s consider some basic numbers, which will help illustrate the relevant concepts. We want to note that our focus is on presenting a rough sizing, and not on providing precise calculations.  Data storage requirements on the desktop are driven by operator activity. Twenty minutes of work on this white paper generated only 25 KBytes on the disk, and most of it relates to word processor formatting. The text-only version spans a miniscule 2 KBytes. The operator-related read rates are not much faster than typing and therefore the “consumption” of the Hard Drive capacity is still rather slow. Watching a movie will push the consumption of the disk capacity to about 1 MByte/second, which is 48,000 times faster than writing a novel. One second of streaming video equates to two days of continuous writing effort.

But even at this supercharged rate, 250 GBytes of capacity is beyond the real needs of a typical desktop user. Reading streaming video would take 250,000 seconds to exhaust the capacity of the drive. This represents nearly 70 hours of pure video content, far exceeding typical desktop application requirements. We have found that over the last few years the hard drive capacity has outlived the usefulness of a system. 

The above gives a good perspective on the practical capacity demands of the desktop for various applications. Many of these numbers are highly applicable for the enterprise environments.

Impact of Capacity on RAID Operation

Average hard drive data transfer speed at the interface is in most cases limited by either the read/write head speed, evident during streaming media processing, or by the average seek time for copies of large file structures. Typically these two numbers are around 50 MBytes/sec for streaming data (sequential access) and 10 MBytes/sec or less when random seeks are required. This means that the shortest time to replace the contents of a 250 GByte hard drive is at least 5,000 seconds, or 1 hour 23 minutes but typically closer to five times this number, or around 7 hours.

Now consider a large enterprise storage system where data integrity is a must, and availability is critical for business continuity. What happens in the situation of a critical failure of a large capacity hard drive? Under the simplest RAID disk configuration, RAID level 1, restoring the redundant data set takes at minimum a couple of hours at full speed of the data transfer, assuming no data path limitations other than the read/write channel. This time is completely lost from the standpoint of data availability and business continuity. If we need to preserve the operation during the recovery time after a failure, the price paid is an extended period of reduced system performance.

Faster Recovery through Better Granularity

Traditionally, the performance of enterprise storage systems has been quantified by a number of “spindles” (or otherwise, hard drive units) available in the system. The spindle count is usually increased in order to reduce average data access latency time, but the spindle count also has a significant impact on data system recovery time after failure. We believe that storage systems constructed with a large number of smaller hard drives would offer a significant benefit for the user affected by the Hard Drive failure without the penalty of performance and capacity reduction. To put it in simplest terms, the recovery after failure would be more manageable and the storage density and performance under normal operational conditions should be comparable or higher. The higher price tag associated with the smaller Hard Drive implementation is well offset by the benefits state above.

In comparison to the RAID array of four 250 GByte drives, a configuration using three arrays of four 80 GByte drives provides comparable storage capacity but significantly reduces the time required to rebuild the system after drive failure. Rebuilding one-third of the amount of data, even at twice slower rate using slower, low-cost drives will take about two-third as much time as before. This much is intuitive and self-evident, but an additional advantage is gained in the availability of the system during rebuild. Since the system distributes data evenly over the three separate arrays, and only the array with the failed drive operates at reduced performance the overall system performance impact is one-third as great for two-thirds of the time of the comparable large-drive array. Ultimately, in our estimation, the real world impact of a single drive failure is 4 to 5 times smaller (22%).

Such systems have not been practical up to now because the parallel hard drive interfaces of SCSI and IDE were unmanageable in high-density systems. Here is where Serial ATA technology presents a very appealing advantage over those legacy interfaces. It provides a fast and narrow data path conducive to switching or multiplexing technologies, and includes power management and hot-swap elements embedded into the interface. (The upcoming serial attached SCSI (SAS) standard attempts to leverage the same qualities but would be at a disadvantage in this situation due to higher cost and more complex implementation requirements, and would still be speed-limited by the read/write channel of the hard drive.)

Construction of a very functional and versatile data storage system using a large number of smaller hard drives is now a practical reality using SATA interface technology. This approach can also compete with large-drive implementations on the basis of physical size when constructed with 2.5” form factor hard drives. The volumetric storage density of these mobile hard drives is close to that of 3.5” drives, and one can argue that implementers can increase system performance despite the slightly lower data rates because of the multiple spindles. While it is necessary to consider the comparable reliability of 2.5” drives, it is undisputable that recovery time after a single drive failure can be reduced by a factor of the capacity ratio with a small correction for speed difference.

The Trend

We believe many hard drive manufacturers are working feverishly on a new and high reliability generation of Serial ATA hard drives compliant with 2.5” and 1.8” form factors. Some of these drives will inevitably find their way into the enterprise environment benefiting those who are reluctant to invest in, or willing to abandon, a high cost SCSI or FC infrastructure. 

We also believe that high end, high performance data storage systems will continue to employ high speed “Enterprise Class” standard form factor hard drives with reduced capacity and elevated spindle speeds, and that SATA will make strong inroads into this market as well.


Additional Resources

http://www-2.cs.cmu.edu/~garth/RAIDpaper?Patterson88.pdf
http://www.acnc.com/04_01_00.html
http://linas.org/linux/raid.html
http://www.webopedia.com/TERM/R/RAID.html