SATA Takes on SCSI update from October 2005

SCSI: The Grizzled Veteran

Since the late 1980s, SCSI has dominated the high-performance storage market for workstations and servers. At the heart of the SCSI controller is a dedicated ASIC that manages all communications with its attached devices. This means a SCSI controller and the devices connected to its bus can function almost independently from the host system. As a result, disk-intensive applications using IDE/ATA in the early 1990s could have required 10 to 15 times the CPU resources used by SCSI under the same circumstances.

Although SCSI's bandwidth and flexibility have improved over the years, the modern Ultra320 SCSI interface is still based on the original parallel-bus architecture, where all the devices on a channel share a common ribbon cable that is terminated after the last device. The Ultra320 SCSI interface supports a transfer rate of up to 320 MB per second and offers a full range of server-class controller cards and disk drives. SCSI hard drives are available in capacities of up to 300 GB, and the smaller, highest-performance SCSI drives operate at a mind-boggling 15,000 RPM. For the small RAID arrays needed for economical servers and workstations, SCSI had no serious competition until the appearance of quality SATA drives and controllers over the past few years.

SATA: The Contender

Serial ATA was developed as the logical successor to IDE/ATA, which arose in the mid-1980s as an alternative to controller-dependent MFM (Modified Frequency Modulation) and RLL (Run-Length Limited) drive systems. Like SCSI, the IDE/ATA interface was based on a parallel connection, but that's where the similarities ended. IDE/ATA was limited to two devices per channel, configured in a master-slave relationship. Rather than being an independent bus controlled by a dedicated ASIC, IDE/ATA depended solely on the system's CPU to handle storage requests. As a result, disk-intensive applications demanded substantial system resources--for some tasks, as much as 90 percent CPU utilization compared with SCSI's 5 percent.IDE/ATA made up for this difference by being affordable. Drives were a fraction of the cost of comparable SCSI disks, and there was no added expense for a dedicated SCSI controller card. When the faster Ultra ATA standard increased disk performance, higher CPU speeds minimized the processing burden, and IDE/ATA storage capacities began matching those of SCSI. This prompted server makers to offer IDE/ATA drives as a low-cost, entry-level alternative to SCSI.

Unfortunately, IDE/ATA has hit its theoretical performance limits, and the electrical constraints of the parallel-bus architecture have made it technically difficult to increase the maximum transfer rate beyond its current 133 MB per second. The new SATA platform has taken advantage of advancements in serial technology. Improvements over parallel IDE/ATA include reduced cable size, CRC error checking for both data and control commands, hot-swap capabilities and an ultrahigh-speed embedded clocking strategy that now supports point-to-point data transfer rates of up to 150 MB per second for each SATA device in the system.

Instead of multiple drives sharing one ribbon cable, each device is connected to a dedicated port on the SATA controller using a thin cable up to 30 inches long. Each connection serves as an independent data bus, so there's no contention between devices and they're not forced to share the interfaces' bandwidth. More important, even though SATA is not hardware-compatible with legacy IDE/ATA drives, the SATA standard is fully compliant with the ATA protocol and compatible with all existing software.

The SATA interface has gained acceptance in the desktop market, with many motherboard manufacturers offering on-board support for both single-drive and RAID 0 or RAID 1 configurations. SATA became interesting for enterprise use when controller manufacturers started coming out with advanced RAID controller cards, in SCSI or SATA, offering virtually identical processing and array-management features. These cards can also off-load most array functions, and therefore offer the same reduced CPU utilization requirements usually associated with SCSI controllers.

There's a broad spectrum of choices when it comes to RAID controllers for both the SATA and SCSI platforms. We established a baseline set of features that would be appropriate for enterprise use, then chose economical SATA and SCSI controllers that fit this set. We required each controller to support a four-drive RAID 5 array with a fifth drive as a hot spare. The controller also would have to offer 64-bit PCI compatibility, online capacity expansion, battery-backed write caching and support for SNMP, as well as the new SMART (Self-Monitoring Analysis and Reporting Technology).The LSI Logic MegaRAID SATA 150-6 controller and the Adaptec 2130SLP Ultra320 SCSI controller offer all these features, supporting JBOD volumes and RAID 0, 1, 5, 10 and 50. For hard drives, we approached Western Digital, the only vendor of 10,000-RPM, enterprise-class SATA drives, which provided five 74-GB Raptor units. On the SCSI side, Seagate sent samples of its 10,000- and 15,000-RPM enterprise-class Cheetah drives for comparison.

In our Real-World Labs® in Green Bay, Wis., we set up a dual-Xeon processor, Dell PowerEdge 2650 server as a test bench. Aside from the different cabling strategies, the process of building an array was virtually identical for both SATA and SCSI. We configured the devices at the BIOS level, through simple menu-driven interfaces for configuring the adapter and specifying the tasks for the individual attached drives. We built each array by assigning the hot spare, selecting and initializing the drives, and specifying the RAID level, stripe size, caching scheme and name for the array. Both controllers also offered useful application software for monitoring and reconfiguring the array at OS level and provided support for error notifications.

In configuring each controller, we kept most of the factory defaults but turned off write caching to get a more accurate assessment of the hardware's true write capabilities. Some admins avoid write caching for fear of data loss in the event of a power outage. Under normal circumstances, however, we would leave write caching on to benefit from the substantial write-performance gain--especially if the cache were battery-protected. Each controller has a battery backup option that can sustain the contents of the write cache for up to 72 hours, making it possible to write the data to disk as soon as power is restored.

We also disabled the background RAID construction and verification option, which lets the array become active immediately but causes the system to operate at a much lower performance level while the array continues to build in the background. This is a useful feature in busy environments, where it's difficult to schedule downtime to completely build an array, but could affect our benchmark accuracy.

With each of our arrays fully configured, the 10,000-RPM SATA system was clearly the least expensive setup, at a street price of $1,439, with the 10,000-RPM SCSI combination coming in at $2,204 and the 15,000-RPM SCSI solution priced at $3,554. The relative performance of these configurations turned out to be somewhat less defined, however, with SATA presenting fascinating statistics in some of our tests (see our pricing chart right.)The results from our Iometer tests (see "How We Tested SCSI and SATA Drives," below) highlighted SCSI's performance edge when it comes to larger, 1-MB file transfers. The Seagate Cheetah SCSI drive arrays showed maximum throughputs of 223.4 MB per second (10,000 RPM) and 260.9 MB per second (15,000-RPM); the 10,000-RPM Western Digital Raptor drive array's maximum throughput came close to the higher-RPM Cheetah's score, at 153.9 MB per second. The real surprise came in the 64-KB read tests, where SATA dominated the field with an impressive 117.2 MB per second, against the 10,000-RPM SCSI unit's 65.5 MB per second and the 15,000-RPM model's 84.2 MB per second performance. As expected, our read tests for 64-KB and 1-MB files fell between 12 percent and 15 percent of the write performance and proportionally followed each array's read capabilities. Raw-disk write performance is always considerably slower than reads, explaining the popularity of large write caches. Enabling the write cache on a controller can increase write performance twentyfold or more, depending on cache and transfer size.

Figure 1:

Given its lower cost, matching functionality and reasonable performance--especially at smaller transfer sizes--it's easy to envision a place for a SATA RAID array in both economical entry-level servers and general-computing workstations, provided you build the array using high-speed, enterprise-class SATA drives. For applications requiring high throughput, such as video editing, or for handling the massive transaction requirements of a busy database, SCSI clearly remains the better performer and therefore the better choice.

The SATA platform has another limitation right now: a lack of drive choices for the enterprise platform. The vast majority of SATA drives are being manufactured for desktop-level use and don't offer the higher rotational speeds and duty cycle of SCSI drives. Although it's true that SATA drives are available with up to 500 GB of storage capacity, only Western Digital sells 7,200- and 10,000-RPM enterprise-class drives, with maximum capacities of 250 GB and 74 GB, respectively. This limited availability shouldn't present a major problem for economy-minded buyers, though it may be an issue for large-scale installations.

Two Platforms on the MoveThe SCSI interface, like IDE/ATA, is a victim of the same limitations of the parallel bus; and though there is an existing standard for Ultra640 SCSI, it's entirely possible that Ultra320 will be the last iteration of SCSI as we know it. Parallel SCSI's proposed replacement, SAS (Serial-Attached SCSI), will arrive this year, and the SAS standard will share a number of similarities with the SATA platform.

SCSI's familiar ribbon cables, terminators and ID issues will soon be history. In fact, SATA and SAS share the same backplane design, so you'll eventually be able to incorporate both drive types in a mixed array that meets your performance requirements. SCSI will also grow because the underlying protocol is incorporated into a number of major industry standards, including iSCSI, SCSI over ST and the SCSI RDMA (InfiniBand) protocol.

The SATA platform isn't standing still either. Work on SATA2 is under way and will introduce port sharing, increase the maximum transfer rate to 300 MB per second, and implement native command queuing to let the controller reschedule active drive requests and improve data-throughput efficiency. Be that as it may, SATA will still be regarded as the low-cost, low-performance alternative to parallel or Serial SCSI.

These technologies have been designed for convergence from the start, and you're certain to find future RAID controllers combining SATA and SAS on the same card. However, like SCSI and IDE/ATA, future SAS development will remain one step ahead of SATA in performance and scalability, in part because of SCSI's focus on the enterprise. And given that focus, storage manufacturers will continue introducing their highest-performance drives on the SCSI/SAS platform.

Steven Hill owns and operates ToneCurve Technology, a digital imaging consulting company. Write to him at [email protected].Look out, SCSI. SATA is making its way into the data center. Once the domain of cheap PCs, IDE (Integrated Drive Electronics) and ATA (Advanced Technology Attachment) drives have moved from the parallel to the serial bus, leaving behind such drawbacks as heavy CPU utilization and slow data transfer. Today's Serial-ATA, or SATA, drives have smaller cables than their predecessors, perform CRC error-checking, are hot-swappable and can transfer data at rates of 150 MB per second. With the appropriate RAID controller, pressure is taken off the CPU. Best of all, these drives are far less expensive than their SCSI counterparts.

But does SATA really perform as well as SCSI hard disks in an enterprise data center? To find out, we tested a setup with a RAID 5 array of five 74-GB, 10,000-RPM Western Digital Raptor SATA hard disks, attached to an LSI Logic MegaRAID SATA 150-6 controller. We compared that setup with two SCSI arrays, using one set of 10,000-RPM Seagate Cheetah drives and a second set of 15,000-RPM Cheetahs. Both SCSI setups used an Adaptec 2130SLP Ultra320 SCSI controller. Although SCSI dominated some of the tests, SATA did extremely well in small-file transfers and would be well-suited to an entry-level server environment, especially where cost is a concern. Our SATA array's street price, $1,439, was well below either of our SCSI setups' prices: $2,204 for the 10,000-RPM drive-based array and $3,554 for the 15,000-RPM drive-based array.

We used a dual-processor Dell PowerEdge 2650 Server with 1 GB of RAM. All tests were based on a four-drive array with a fifth as a hot spare. The SATA RAID 5 array was built using the adapter's optimal 64-KB striping, and both the 10,000 and 15,000 RPM Ultra320 SCSI RAID 5 arrays were built using the adapter's optimal 256-KB striping. For benchmarking, we used the Iometer test application, version 2004.07.30. All tests allowed one outstanding I/O operation and were set to run for 1 minute with a 20-second ramp time.

Figure 2:

For our 64-KB and 1-MB linear read/write tests, we used four workers on one target and set read/write distribution to 100 percent read and 100 percent write. Random/sequential distribution was set at 100 percent sequential. Our IOps test used a 512-byte transfer-request size and was run with four workers on one target at 100 percent read and 100 percent sequential.The NWC Custom Test was set to transfer-request sizes of 8 KB with 33 percent access distribution; 8 KB, with 34 percent; and 64 KB, with 33 percent. On the first 8-KB segment, we set read/write distribution to 100 percent read, and random/sequential distribution to 100 percent sequential. On the second 8-KB segment, we set read/write to 67 percent read and 33 percent write at 100 percent random. On the 64-KB segment, read/write was 100 percent read at 100 percent sequential.

All Storage Pipeline product reviews are conducted by current or former IT professionals in our Real-World Labs® or partner labs, according to our own test criteria. Vendor involvement is limited to assistance in configuration and troubleshooting. Storage Pipeline schedules reviews based solely on our editorial judgment of reader needs, and we conduct tests and publish results without vendor influence.

Steven Hill owns and operates ToneCurve Technology, a digital imaging consulting company. Write to him at [email protected]