Test driving a PCI-based DLT library
As Windows 2000 invades the enterprise data center, sites will have to confront the explosive growth in disk data. Further complicating this scenario is the distributed nature of the BackOffice family, which often has enterprise applications such as Exchange and SQL Server running on separate servers. Automated tape libraries that can service multiple servers either through a direct connect or a smart software scheme will be essential.
By Keith Walls
Tape library robotics may not be as entertaining as a Hollywood high-tech monster movie, but they do offer a certain gratification as they swing into action--keeping track of the identity of each tape and tracking where it sits in the library. That satisfaction is intensified by the knowledge that the only option would be to get in there and change the tapes manually, with all of the tracking and matching of tape labels and their destinations falling into the hands of a human operator.
Once we decide to commit data to a tape library, we need to be able to find the individual cartridge containing the data, and we need to be able to find it indexed by content or tape label. As a result, reliability of the machinery and reliability of its detailed operation are important.
We must also be able to remove tape cartridges and groups of tape cartridges easily, accurately, and quickly so that we can satisfy the off-site storage and tape rotation procedures that support the disaster-recovery plan for the organization. Since tape management plays a significant role in any data-protection plan, CTO Labs examined several practical aspects of deploying a tape library in the context of a Windows NT enterprise operation. We then created a requirements plan for deploying any tape library.
One of our principal goals was to be able to issue a command sequence such as:
1. Load next scratch tape in a drive
2. Initialize tape in drive
3. Perform backup to this drive
4. If the backup overruns the current tape then,
4.1. Unload current tape
4.2. Move current tape to offsite bin
4.3. Load next scratch tape into a drive
4.4. Initialize tape in drive
4.5. Continue backup operation
5. Unload current tape in the current drive
6. Move tape cartridge to offsite storage bin
CTO Labs tested the P1000 tape library from ATL, a subsidiary of Quantum. It was configured with four DLT7000 tape drives. This tape library presents all the hardware capabilities needed to fulfill our stringent requirements in a practical and affordable configuration. The library holds a total of 30 DLT7000 cartridges. Each cartridge holds 35GB with no data compression and 70GB with a compression ratio of 2:1. In examining tape backups at our site, CTO Labs typically achieves a slightly lower compression ratio, on the order of 1.6:1. That puts the maximum storage capacity of the P1000 tape library on the order of 1.7TB to 2.1TB.
Configuration for performance
The P1000 can support from one to four drives, so the most demanding part of the site preparation for configuring the ATL Library was making a decision on how to configure the unit across the available differential or single-ended SCSI buses. Because of the distances involved in setting up a library, differential SCSI was the default for CTO Labs.
The P1000 library can be installed to take into account all of the practical configurations that might be needed. A fully configured library presents a total of five SCSI devices: the tape changer plus four tape drives. For each tape drive, the maximum throughput of normally compressible data (at a compression ratio of 1.6:1) is approximately 7MBps. Theoretically, an Ultra SCSI rail should perform at 40MBps. Nonetheless, we must subtract some of the throughput capability of each SCSI bus to account for data and command collisions for the multiple devices. That would set the expectation that each SCSI rail would support nearly three DLT7000 drives. However, CTO Labs has found that a more conservative two drives per SCSI rail is a much safer bet for optimal performance.
We configured the ATL P1000 with a single four-way Dell PowerEdge 6100 server via two Adaptec 2944UW differential SCSI controllers. Nonetheless, the library can potentially be configured on five different machines. The tape changer, which has its own SCSI connection, maintains control of the entire tape library independent of the SCSI ownership or distribution of the tape drives.
If you think that capability hints at the creation of a server area network, you are absolutely correct. In fact, ATL will soon offer the option of installing a Fibre Channel interface on the P1000. The presence of an internal PCI bus in the library is the secret to the unit`s performance.
The arrangement of tapes in the ATL tape library is also of interest. At the front of the cabinet are two removable bays that carry eight DLT cartridges each. Immediately behind the removable bays is the robot mechanism. Behind the robot are the four DLT7000 tape drives and another rack for tape cartridges.
In normal operation, the front doors to the ATL P1000 are electromagnetically locked. The locked doors prevent access to the removable bins. In order to remove either of the eight-cartridge bays, an operator must enter a numeric password on the library`s front control panel and LCD before being able to unlock the doors manually. Access to the operator security level allows removal of tapes, removal of the transport bays, and tape movement and manipulation.
The control panel provides menu-drive capability for operators to move tapes, load tapes, and unload tapes. The display is a touch-sensitive screen whose contents closely resemble a GUI. The front-panel control should be useful in rack-mounted units; however, our library was clearly designed to stay at floor level, on its wheels. As a result, we spent a lot of time sitting on the floor in order to read the LCD and eliminate parallax errors.
Nonetheless, the library`s front panel provides excellent control and a full set of functions for managing and servicing the unit. The capabilities of the ATL front panel are unsurpassed, particularly when compared with units that have a small number of buttons and lights. Such units generally use a different flash duration to indicate different conditions. Deciphering the light flashing patterns on these devices is confusing and requires that the manual be kept close at hand.
The ATL main screen presents several tabs that resemble the tabs of property pages. From the main screen, the user can select overview, operator, and service menus. There is also a service level of security, where the library`s configuration can be changed and diagnostic tests performed.
The library`s two removable front bays hold a total of 16 tapes. By opening the front of the cabinet, the bays, complete with the tape cartridges they hold, can be removed. In addition to the two eight-cartridge removable bays, there is a single-cartridge loader. An operator or a controlling program can eject or insert a single cartridge as required. However, the tape cartridges in the rear of the compartment are not accessible, unless you remove the cover of the cabinet.
This requires a significant degree of planning when using the ATL to rotate tapes through off-site or shelf storage. All the tape cartridges that must be stored outside the box must first be moved to the front bays. If more than 16 cartridges are to be removed, more than one pass through the unload-reload loop is required.
Controlling the changer
The ATL P1000 is quite capable in terms of both performance and function. The robotic arm is fast enough to service all four tape drives. The two removable storage bins are easily adequate for most purposes. Nonetheless, to harness the full potential of the ATL P1000, especially if it is to be configured with multiple distributed servers, requires some intelligent software.
Many existing software suites provide at least some of the functions that are needed to fully utilize the P1000`s capabilities for automation and unattended performance. For example, the P1000 has the ability to separate, categorize, and group cartridge bins. It would be convenient to be able to define slots 1 through 14 as the "storage bins," slots 15 through 22 as "removable bin #1," and slots 23 through 30 as "removable bin #2."
Most packages offer a way to group tape cartridges and even to group media elements. Nonetheless, we have not yet seen the capability to easily express an operations construct such as, "move a tape to the next open slot in the removable bins."
Of even greater complexity is the task of switching complete sets of tapes. Rotating tapes is a natural component of any media management scheme. Suppose that 16 new tapes are to be inserted into the library using the removable storage bins. That night, you would like to use existing tapes in the library to perform an incremental backup (record only data that has changed). When that backup is complete, all those tapes are to be placed in the output removable bins, and the new tape set is to be used to perform a full backup of data, which is to be left in the library. With only five available slots (the single-cartridge loader and the four tape drives), the problem quickly resembles a shell game.
Another issue is that the server to which the changer is attached represents a single point of failure. If the machine controlling the changer goes down, tapes can only be moved either manually or by issuing commands at the display console. Currently there is no provision in MSCS to share a changer, so it is not possible to fail the changer over to an alternate MSCS node in the event of failure.
Numbers of numbers
As CTO Labs presented in our earlier review of a standalone Quantum DLT7000 drive, (see "64KB backup blasters," BackOffice, Sept. 1997, page 17) the rate at which data is transferred to tape is dependent on the compression ratio the drive is able to achieve with the data presented to it. With no compression, the data rate of the DLT7000 peaks at 5MBps (300MB per minute). At the average compression ratio of 1.6:1 that we consistently encounter at CTO Labs, the average data rate is between 7MBps and 8MBps (about 450MB per minute).
This is exactly what we saw with the drives in the ATL library. Individually, they behaved just as our desktop Quantum unit had performed. Throughput from a single drive peaked at about 450MB per minute with both our tape benchmark, which writes a stream of data from memory, and in performing backups with CA ARCserve.
Far more important, however, is the ATL unit`s ability to scale up in throughput. This is a result of the unit`s Prism Architecture, which puts network interface cards and I/O controllers on an internal PCI bus. With two of the library`s drives connected to the Dell PE6100 via either one or two SCSI busses (Adaptec 2944UW cards), throughput virtually doubled. Even more remarkably, with four drives spread over two SCSI busses, throughput on the CTO Labs tape benchmark soared to 1.7GB per minute.
With ARCserve and data coming from RAID10 SCSI disks, we had to settle for a less stratospheric .85GB per minute performance. Nonetheless, that pegs real-world backup performance with our configuration at about 50GB per hour. On a eight-hour shift, that would enable the backup of 400GB of data using the eight cartridges in one of the two removable storage bays.
Looking ahead to the Removable Storage Manager in Windows 2000, and the Fibre Channel interface coming to the ATL P1000, we hope to be able to shortly test software that will manage the movement of tape in synchronization with backup operations across several machines.
NOTE: This article is reprinted with permission from BackOfficeCTO magazine, a sister publication of INFOSTAR. For more information or to subscribe, visit www.backofficemag.com
Running multiple jobs with CA-ARCserve, we were able to back up data at the rate of .85GB per minute. In this case, all of the data for each of the simultaneous backup jobs resided on a single RAID10 volume set.
Connected to two Adaptec 2944UW cards on a Dell PowerEdge 6100 server, the library handled the CTO Labs tape benchmark with perfect linearity. With four drives, we were able to write data at the rate of 1.7GB per minute.