Our lab combines Quantum's Super DLT drives, Overland Data's Neo Series libraries, and BRU-Pro backup software. The results follow.
By Jack Fegreus
Today, multiple choices in both linear and helical-scan formats offer users more freedom than ever to choose the tape option that best meets their requirements. For the current generation of enterprise-class tape systems, the battle lines are drawn among Super DLT (developed by Quantum), Linear-Tape Open (LTO) Ultrium (backed by Hewlett-Packard, IBM, and Seagate), and Mammoth-2 (from Exabyte). While all three technologies offer extraordinary throughput speed and cartridge capacity, the explosive growth of storage has driven many midrange enterprises and even large departments that only a few years ago would never have considered tape automation into the market for a tape library.
Quantum's Super DLT and LTO Ultrium drives from HP, IBM, and Seagate continue the fundamental difference between linear tape technologies and helical scan as currently represented by Exabyte's 8mm Mammoth-2 line. To drive throughput, linear tape drives move tape rapidly past stationary heads at a speed of up to 160 inches per second (ips). On the other hand, helical-scan drives rely on a slow-moving tape-1.8ips-crossing a fast spinning set of heads mounted on a drum or scanner. The net result is a relative head-to-tape speed for the Mammoth-2 of 547ips.
The necessity to move the tape rapidly and stop and reverse direction to record on the next set of tracks places a high degree of stress on the tape media, which makes the tape more susceptible to wear and damage. In the case of current DLT drives, there can be as much as 133 grams of tension on the tape, which is more than an order of magnitude greater than the Mammoth-2's servo-controlled, direct-drive dual-reel mechanism, which only exerts 12 grams of tension on the tape.
This fundamental difference naturally results in distinct operating characteristics. Super DLT and LTO write data in long parallel tracks that run the length of the tape. Filling a tape with data requires numerous passes as the tracks wind in a serpentine fashion. Helical-scan technology, however, writes data in short angled tracks that run across the width of the tape. Since the axis of rotation is not orthogonal to the tape's line of motion, the tracks are angled across the width of the tape, and all of the tape is used on just one pass.
Implications for performance
These differences in the way data is laid out on the tape raise interesting implications for performance. The universal truth for all tape drives is that performance is totally dependent upon the ability to keep the tape streaming across the head. Anything that interrupts the flow of data will significantly impact overall performance as the system is forced to stop and reposition the tape. When such an event occurs, the physical rather than relative speed at which the tape is moving past the head will be a more determining factor in how long it takes to reposition the tape and resume writing data.
To pack 110GB of uncompressed data on a single Super DLT cartridge-the previous generation DLT8000 has a native capacity of only 40GB-Super DLTtapes are formatted with 448 data tracks, with each track written at a density of 133Kb per inch. That's an areal density of 166Mb per inch2. With the tape streaming by the heads at 160ips, Quantum faced serious technological hurdles to ensure a low-error bit rate as well as the ability to scale the areal density over the coming years by more than two orders of magnitude. To this end, Quantum pioneered a radical new technology for the Super DLT drive dubbed Laser Guided Magnetic Recording (LGMR). For the first time in tape history, Quantum joined laser technology with magnetic technology in a single tape system.
LGMR ensures higher cartridge capacities by servo-ing from optical targets on the back side of the media. As a result, 100% of the magnetic surface, as well as 100% of the magnetic heads, is dedicated to reading and recording data tracks. Traditional magnetic tape designs reserve 10% to 20% of the recording surface for storing servo track information. In contrast, Quantum laser-etches optically readable servo tracks on a specially formulated back coating of the media and uses a three-beam hologram configuration for exact tracking.
In addition, the laser servo tracks cannot be magnetically erased. This indelible servo information eliminates the need to magnetically pre-format tapes and makes it possible to bulk erase Super DLT cartridges. Traditional media must have the servo tracks re-recorded after a bulk erase-a process that is highly susceptible to environmental variables and, therefore, highly discouraged.
To handle the higher areal bit density of Super DLTtape, Quantum also introduced magneto-resistive cluster (MRC) read/write heads. The heads deliver higher data-transfer rates and are less susceptible to environmental conditions such as temperature and humidity than traditional heads of equal size.
In addition, the Super DLTtape drive, like the Ultrium LTO and Mammoth-2, implements advanced partial response maximum likelihood (PRML) channel technology, which is used by many hard disk drive manufacturers. In essence, the channel compares the measured signal from the tape with a known waveform in order to interpret the data. PRML attempts to correctly interpret even small changes in the analog signal, whereas peak detection relies on fixed thresholds. As a result, a drive using PRML can correctly decode weaker signals and read/write data at a higher bit density.
The net result is a very fast tape drive, which is highly dependent on the ability to stream very large blocks of data to keep it from pausing and repositioning the tape. We first calibrated the SDLT drives using the obltape v1.0 benchmark at block sizes of 128KB.
Our tape benchmark generates two very different types of data streams: purely random data, and data that falls into a pre-set frequency pattern. The patterned data stream was originally devised and calibrated using Exabyte Mammoth-1 and Quantum DLT 7000 tape drives, which implemented the Digital Liv Zempel (DLZ) compression algorithm in hardware. This algorithm purportedly provided a 2:1 compression ratio on normal data, so we devised a means of generating patterned data that consistently produced a compression ratio on the order of 1.9:1 to 2.1:1 on those devices.
The obltape benchmark first allocates a large block of memory from which it then streams either patterned or random data to the device. By streaming data directly from memory, the benchmark eliminates bus bandwidth contention with other devices. The data can be streamed in block sizes of 2nKB, where n ranges from 0 to 8.
Using a QLogic Ultra160 host bus adapter (HBA) to connect the library via LVD SCSI, we pegged base uncompressed throughput at 128KB blocks to be 10.7MBps. This level of performance is actually slightly higher than the drive's rather conservative 10MBps rating. That puts the SDLT in line with the Mammoth-2, but trailing the Ultrium LTO drive from HP, which delivered a native throughput of 13.9MBps.
When we ran the compressible data stream with hardware compression enabled, throughput on the Super DLT drive, which implements a DLZ algorithm, doubled as expected to 20.4MBps. In contrast, the Ultrium LTO and Mammoth-2 drives use a new Adaptive Lossless Data Compression (ALDC) algorithm that purports to provide an average compression ratio of 2.5:1 across multiple data types.
In our benchmark, data compression on the HP Ultrium LTO drive was pegged at 2.3:1. In addition, the HP Ultrium implements what HP calls "smart data compression." HP's compression circuitry has a "pass-thru mode" which switches off compression for non-compressible data, which is typical of jpeg and zip files. According to HP, this "pass-thru mode" can be at least 10% more efficient.
This is demonstrated in the worst-case test scenario. When purely random data is sent to a drive while hardware compression is on, the drive attempts to compress the data and wastes embedded CPU cycles, and throughput degrades to less than the native streaming transfer rate as buffer management becomes problematic. As a result of its pass-thru mode implementation, the HP Ultrium LTO drive showed the least variance in performance-dropping from 13.9MBps to 13.7MBps when purely random data was sent to the drive with hardware compression turned on.
As data becomes more mission-critical, and downtime becomes more costly, IT site managers are requiring high-availability features on tape automation devices. So while the Super DLT drives provide a technology-rich basis for a tape subsystem, the real magic of the Overland Data Neo Series libraries lies in the automation robotics. In fact, Overland is neutral when it comes to drive technology and provides robotics for LTO, Sony AIT-2, and other drive technologies.
Each 5U-high Neo Series library module supports up to two drives and 26 media slots, including a mail slot. Along with power supplies, controllers, and robotics for high availability, the library uses "hot-pluggable" drive carriers, which allow drive replacement without interrupting backup-and-restore functions. These hot-pluggable drive trays also provide an easy upgrade path toward future Super DLT drive technologies.
Nonetheless, what separates the Neo Series from some other tape libraries is the ability to scale with multiple units. Up to eight Neo Series libraries can be linked together into a single logical "virtual library." As a result, a multi-module virtual library can be configured in an industry-standard rack with up to 16 drives and 208 media slots. We created a more modest two-unit virtual library in our test scenario, with each module hosting one SDLT drive.
Modules are linked using the XpressChannel, which adds an elevator mechanism for tape cartridge movement. This mechanism moves tapes efficiently from module to module, allowing any tape cartridge to be moved to any available drive or media slot in the system. The XpressChannel includes a 10U motor drive assembly for the first two library modules, plus extensions for each additional module installed in the rack.
Perhaps the most significant benefit of Overland data's architecture is the ability of the robotics to continue operating during hardware fault conditions. Borrowing from the construct of a server cluster, one of the Neo Series modules is configured as the master controller in a multi-module configuration. This master controller module has a fail-over mode whereby another module can take over as the standby master controller. Fail-over management can be instituted either locally through the library's front panel touch screen or via the library's Web interface.
We set up our two-module Neo Series library via the front-panel touch screen on our master module and then monitored the system over the Web. To test the library's fail-over mechanism, we powered down one of the master modules and observed as the standby master automatically took ownership of the Web interface. In this process, the SDLT drive in the primary master library module became "grayed out," and the Status Summary reported that the virtual library could not communicate with that drive. In addition, the tape slots in that drive, which were now inaccessible, also disappeared from view.
This same information was delivered via the SCSI interface from the library management module to the library management software that we were running on our host server.
To test the Overland Neo Series in a real-world backup scenario, we used the Tolis Group's BRU-a backup package familiar to many Linux users. We actually used the new BRU Professional Archive Management System (BRU-Pro), which is built on MySQL.
BRU-Pro integrates the original BRU utility with a MySQL database and an easy-to-use GUI in order to eliminate the need for systems administrators to know complicated command line sequences and data flags. The design goal of BRU-Pro was to allow a systems administrator with no special skills above a general working knowledge of the native operating system to perform backup-and-restore tasks on a variety of machines (clients) from a single workstation (control console) writing to a centralized server (tape server).
BRU-Pro automatically scans all of the SCSI devices on the tape server system and lists the appropriate libraries and tape drives. In essence, BRU-Pro relies on the operating system to discover and categorize the devices. If the operating system recognizes the library, then it should be listed in BRU-Pro, which can invoke a screen that contains a full description of the library and a listing of all of the slots that are present and their current tape status. So with no effort on our part, BRU-Pro came up recognizing both the Overland virtual library and the Super DLT drives without a hitch.
With the new GUI and underlying MySQL database, a systems administrator can now easily configure a backup task with a few mouse clicks. Files can be selected by selecting entire client machines or directories or by expanding the directory tree and choosing specific files.
To ensure the integrity of a backup archive, BRU defaults to an "Auto matically Verify" option that triggers the execution of BRU-Pro's checksum verification program immediately at the completion of a backup. A systems administrator can also scan an archive manually at any time. In addition, a completion report can be automatically sent to a specific e-mail address once a backup has finished.
The enhanced archive cataloging that MySQL brings to BRU is most evident in critical recovery situations. Based on the ID of the user logged into BRU, the Restore menu shows a listing of all of the tape servers and client workgroups that have backup archives that the current user can access. Selecting one of these archives then displays a directory tree from which the systems administrator can choose directories or specific files to restore.
The final phase of our testing involved a series of backups and restores performed with BRU-Pro. In this phase, one of the legacy limitations in the BRU backup engine became clear. To simplify device configuration in the past, BRU was designed with a hard-coded 32KB buffer for data transfers. While this was adequate for tape drives two years ago, today's high-speed drives require data buffers that are 64KB or higher. As the results of our obltape benchmark indicate, with a 32KB buffer, top-end performance results suffer across the board.
BRU was able to keep the drives only as busy as the limitations of a 32KB buffer would permit. Nonetheless, backup throughput of 15MBps is not shabby. Furthermore, the simplicity of configuration and ease of use demonstrated by BRU make up for the temporary performance penalty which, like certain legacy file-size restrictions, are being eliminated by the Tolis Group.
InfoStor Labs scenario
- Super DLTtape drives
- Modular automated tape library with hardware fail-over capabilities
- MySQL-based backup software
What we tested
- Two Overland Data Neo Series LXN2000 automated tape libraries (OEMed by Compaq as the MSL5620SL) (www.overlanddata.com) and (www.compaq.com)
- Two Quantum Super DLTtape drives (www.quantum.com)
- BRU Professional Archive Management System software (www.bru.com)
How we tested
- Red Hat Linux v7.1 (www.redhat.com)
- QLogic QLA12160 HBA (www.qlogic.com)
- obltape v1.0 benchmark
- BRU-Pro easily managed the Overland library and Super DLT drives.
- BRU-Pro throughput performance is currently limited by a hard-coded data-transfer buffer size of 32KB.
Jack Fegreus can be contacted at firstname.lastname@example.org.