|
When disk drive manufacturers began introducing thin film magneto-resistive (MR) heads, scarcely could they have guessed the explosive growth in areal density-storage capacity in bits per inch2-this technology would ignite. With an average disk sporting a capacity of 8GB, tape drive manufacturers today face enormous pressure to increase data throughput and cut backup time. To this end, they have adopted key technologies from the disk industry: multiple MR heads working in parallel and Partial Response (PR) recording channels.
Able to produce a high output signal per unit of track width, MR heads are well suited for use in tape drives. Nonetheless, as areal recording densities and transfer rates are pushed higher, noise bandwidth increases and the signal-to-noise ratio drops with each new product generation. In disk drives, Partial Response recording channels are an established means to restore the signal to noise ratio to an acceptable level. Now PR channels have wound their way into departmental- and enterprise-level tape drives capable of handling 8-to-24MB of data per second.
While there is universal agreement on these two technologies, there remains a divergence of opinion on how best to move the tape with its bits of encoded data past the heads to maximize data throughput. The most direct method is to move the tape at very high speed-upwards of160 inches per second (ips)-directly past the heads in a linear fashion. This is the approach taken by the long dominant Digital Linear Tape (DLT) technology from Quantum. This will remain the approach taken by Quantum's next- generation SuperDLT and the new Linear Tape Open (LTO) format from HP, IBM, and Seagate.
Helical scan technology is the alternative to moving tape directly at high speed and is used in Digital Data Storage (DDS)-better known as 4mm DAT- and 8mm tape formats. With helical scan, the tape moves at a more leisurely speed of under 2ips across a very fast spinning drum on which are mounted multiple sets of read/write heads. In the case of Exabyte's next-generation Mammoth-2drive, the spinning heads induce a blistering relative head-to-tape speed of 547ips.
These fundamentally different tape transport schemes result in equally different operating characteristics. Filling a tape with data requires numerous passes with a linear-actually serpentine-format. In the case of the DLT 7000, it takes 52 passes to fill the tape. In a helical scan configuration, the drum's axis of rotation is not orthogonal to the tape's line of motion technology. As a result, data is written in short angled tracks that run across the width of the tape. All of the tape is utilized on just one pass.
More importantly, data throughput on any tape drive-linear or helical scan-is dependent upon continuously streaming the tape over the heads. Any interruption to the flow of data forces the drive to stop and reposition the tape. When such an event occurs, the physical rather than relative speed of the tape determines how much tape will have to be rewound before it can be repositioned and the drive can resume writing. Even the normal stopping and reversal of direction in linear tape to switch tracks places a high degree of stress on the media, leaving it more susceptible to wear and damage.
Nonetheless, the performance of previous-generation helical scan drives lagged significantly behind high-end linear competitors. While, in the number of units sold, DDS technology was the top tape technology in the PC server market, DDS-3's paltry native transfer rate of 1MB per second and uncompressed cartridge capacity of 12GB left much to be desired. Even Sony's AIT and Exabyte's Mammoth drives, with native transfer rates of about 3MB per second, were overshadowed by DLT 7000/8000 drives.
The performance gap disappeared early this year with the introduction by Exabyte and HP of their new generation of helical scan drives, which increased native performance by fourfold and threefold, respectively. The new DDS-4 drive, dubbed the HP DAT40e, sports a native transfer rate of 3MB per second and an uncompressed cartridge capacity of 20GB. The HP specifications put the HP DAT40e far more in line with the backup needs of servers now being sold to small- and medium-sized businesses. Assuming a nominal 2-to-1 data compression ratio, DDS-4 drives and tapes provide the lowest-cost devices and media to back up 40GB of data.
OpenBench Labs benchmarked both an HP DAT40e external DDS-4 drive with a single-ended SCSI-2 interface, as well as an Exabyte Mammoth-2 external drive, which sports a low voltage differential (LVD) Ultra2 SCSI interface. All tests were conducted on a Dell 2400 PowerEdge server running Red Hat Linux 6.2 and Windows 2000 Advanced Server. We tested the HP DAT40e with an Adaptec 2940 SCSI host bus adapter and the Exabyte Mammoth-2 with a QLogic QLA12160 64-bit PCI host bus adapter.
To test both drives, we ran the OpenBench Labs tape benchmark v1.0. This benchmark allocates a large block of memory from which it streams data to the device. By streaming directly from memory, the benchmark eliminates bus bandwidth contention with other devices. The data can be streamed in block sizes of 2n KB, where n ranges from 0 to 8.
Since we found no difference between 64KB and 128KB block writes on Linux, we chose 64KB writes to be compatible with Windows 2000. Interestingly, we found at least one tar-based tape backup utility that defaults to 10KB writes, which provides roughly 40% of the theoretical peak throughput. (OpenBench Labs will be examining this issue in much greater detail in the coming months.)
The tape benchmark generates two types of data stream: purely random data and data that falls into a preset frequency pattern. The purely random data is not compressible and therefore provides a worst-case scenario. The drive will attempt to compress this data, which is typical of highly compressed JPEG, ZIP, and e-mail archives. In so doing, the drive will waste embedded CPU cycles trying to compress incompressible data and throughput will degrade.
The patterned data stream originally was devised and calibrated using Exabyte Mammoth-1 and Quantum DLT 7000 tape drives on Windows NT 4.0. This data stream consistently produces a compression ratio on the order of 1.9-to-2.1 on these devices under Windows NT/2000. These numbers are highly comparable with backup results on real data that is not stored in a compressed format on these systems. With this in mind, we used 3GB data sets in all of our tests in order to reach a repeatable steady state in terms of data compression and throughput.
In all of our tests, the Mammoth-2 drive provided eye-popping performance increases over the first-generation Mammoth drive. The OpenBench Labs benchmark pegged base throughput with no hardware compression at 10.9MB per second. The base throughput of the HP DDS-4 drive was pegged at 2.9MB per second, putting it on a par with 8mm AIT-1 and Mammoth-1 drives.
When we ran the compressible data stream with hardware compression, throughput on the Mammoth-2 soared to 22.8MB per second on Windows 2000 Advanced Server and 26.8MB per second on Red Hat Linux 6.2. Differences between Linux and Windows for uncompressed performance were statistically insignificant. This represents a compression factor of 2.46, which is in line with the stated 2.5:1 average compression ratio for the drive, which employs a new Adaptive Lossless Data Compression (ALDC) algorithm.
In contrast, we measured a compression boost of 2.62X.on the HP DDS-4 drive to 7.6MB per second under Windows 2000 and a spectacular 14.3MB per second for a 4.93X increase under Linux. Interestingly, the HP drive uses the same DCLZ compression algorithm used in the previous generation DDS-3 drives. The measurable boost in the new HP drive comes from improvements in the SCSI firmware and larger buffers in the compression engine chain.
Using a firmware-based "intelligent disconnect," the DDS4-4 drive optimizes the time to burst data at 40MB per second into a mini-buffer which is used to hold the data before the compression engine. An 8MB buffer then holds up to 20 compressed DDS "groups" of data. With its tuned SCSI bandwidth utilization, the HP DAT40e moves data through the compression chain faster. As a result, if SCSI operations are more efficient under Linux, then all of the drives should exhibit gains in compression performance, with the greatest boost going to the HP DAT40e. Not only did we measure this result, but also measured a greater negative impact on the HP DAT40e under Linux when streaming incompressible data with compression turned on.
To get their threefold and fourfold native performance boosts, both HP and Exabyte dramatically increased the rotation speed of their helical scan drums, in order to increase the relative head-to-tape speed. In addition, the read/write preamplifier circuitry is mounted directly on the rotating drum to put it as close to the heads as possible to maintain a good signal-to-noise ratio. This was a particularly tricky feat, as these components on the HP drum have to cope with forces on the order of 2200g as the drum rotates at 11,480rpm.
Placing the signal amplification circuitry for both the read and write heads on the drum results in nearly equal signal strength for both read and write channels. As a result, the heads can be positioned closer together.
On both the HP and Exabyte, the heads are now positioned at 90° intervals rather than 180° intervals around the drum. The big difference is that the DAT drive has 2 read and 2 write heads, while the Mammoth-2 drive doubles the number of heads to 4 read and 4 write heads. On each revolution of the scanner, the Mammoth-2 writes 4 tracks of data while simultaneously reading the previous 4 tracks of data. Thus, by doubling the write time and doubling the write heads, Mammoth-2 raises performance levels by a factor of 4 over the previous generation of Mammoth drives.
Another important technology that each helical scan drives utilizes to improve signal detection is a Partial Response read channel. This technology originated at NASA to read weak signals from satellites deep in space. In essence, the channel compares the measured signal from the tape with a known waveform in order to interpret the data.
In particular, HP starts with a Class 1 Partial Response Maximum Likelihood (PRML) recording technique and enhances the standard two-state Viterbi trellis with a six-state Advanced Sequence Detection (ASD) scheme. Exabyte, on the other hand, is the first to use an enhanced Class IV PRML (EPR4) technique,which is used by many HDD manufacturers. EPR4 increases the number of times and levels at which each magnetic transition is sampled from two to three.
To provide better contact between the tape and the read/write heads on the drum, both HP and Exabyte have machined grooves on their respective scanners to create a negative air bearing surface (ABS), which draws the tape in towards the drum surface. Exabyte has added a conditioning head in front of each read and write head pair. This head acts as an air dam to reduce drag. It lifts the tape before it reaches the read and write heads and ensures that the tape touches both heads with only enough force to reliably read or write data.
Media, however, continues to set the Mammoth family apart from all other tape drives, as it uses Advanced Metal Evaporated (AME) media rather than the metal particle media used in DDS tapes. The magnetic layer of AME media is applied by evaporating the magnetic material onto the base layers in a vacuum chamber using an electron beam. This process creates two magnetic layers of pure cobalt that don't contain any binders or lubricants. These are then coated with a hard carbon protective layer and then a lubricant. The result is a tape with greater magnetic density, reduced physical thickness, and a cartridge capacity of 120-to-150GB with nominal data compression.
OPENBENCH LABS SUMMARY
What we tested
Enhanced Software Technologies BRU v16.0 and XBRU v1.17HP DAT40e external DDS-4 drive Exabyte Mammoth-2 external drive
How we tested
Dell 2400 PowerEdge server running Red Hat Linux 6.2 Adaptec 2940 SCSI host bus adapter for the HP DAT40e QLogic QLA12160 64-bit PCI host bus adapter for Exabyte Mammoth-2
Key findings
· In all cases, data compression under Red Hat Linux 6.2 was far more effective than Windows 2000 Advanced Server.
· Base performance of HP's DDS-4 DAT drive proved on par with
an 8mm Sony AIT 1 and Exabyte Mammoth-1 drives, and data compression under Linux is nothing less than stellar. The HP drive creates a new low-cost standard for departmental servers.
· The Mammoth-2 drive reveals very impressive performance
increases over the first-generation Mammoth drive with data throughput reaching 26.8MB per second.
About our benchmark
OpenBench Labs tape benchmark v1.0 generates two types of data streams: purely random, uncompressible data, and data in a preset frequent pattern. We chose 64KB writes as compatible with Windows 2000.
| 
