Testing HP`s SureStore Optical Jukebox
MO jukebox puts 80GB of data in the space of a small RAID array, but automation software gives IT/IS Labs a bit of the blues.
By Jack Fegreus
Not long ago, the big problem for any corporation was coping with information scarcity. Today, many would gladly trade our current information glut for those halcyon days of yore. Smart business chiefs, however, see this glut as their key to competitive advantage. These leaders are in the vanguard of the move to informationalize business processes, and many now find themselves storing volumes of data once thought of as astronomical.
Dealing with all of the rapidly expanding demands for on-line data storage can become quite an expensive proposition. While the cost of purchasing a hard drive has plummeted precipitously, the ancillary costs that go with maintaining a large disk farm have been on the rise. The tradeoff for higher storage capacity has been faster rotation speed, higher bit density, and lower mean time before failure (MTBF). As a result, data centers face rising environmental costs as drives draw more power, dissipate more heat, and require closer monitoring.
The longstanding answer in the mainframe and UNIX worlds to the problem of expanded on-line storage is the implementation of magneto-optical (MO) jukeboxes for fairly stable data that is much more likely to be read than written. The digitization of paper documents, pictures, maps, X-rays, fingerprints, and other images has long been a staple application for MO media. That`s because the I/O access patterns of imaging applications are dominated by large data transfers of entire files at once.
Widely accepted by high-end IT sites, MO disks are rugged, and removable, and they provide more than 100 years of disk archival life. IT sites typically use MO disks for data that is far too active to be successfully archived and taken off-line. Instead, it has been classified in the nether world of "near-line" storage. Since the same applications used for on-line storage can access near-line storage, optical jukeboxes provide complete storage device transparency. System administrators can now organize and move data to different storage devices without imposing new access methods on users.
Now that Windows NT servers are moving up in class, more sites are requiring the types of high-end storage solutions for on-line RAID, automated backup, and near-line repositories to which they have become accustomed on legacy operating systems. In recognition of this fact, Microsoft has enlisted HighGround Systems to build a snap-in control to manage removable media devices in Windows NT 5.0. This control is in the current interim technical beta (build 1773). But under the current NDA, if we showed it to you, we`d have to kill you.
So what does a BackOffice site do if it is running network applications like document and image management, CAD/CAM, video-on-demand, and desktop publishing today? How do you satisfy the demand for more storage and faster access at a reasonable price right now?
Billions and Billions of Bytes
As we said, the hardware solutions have been around for quite some time. One of the premier vendors, Hewlett Packard, is now introducing drives and cartridges that double the capacity of high-end rewriteable magneto-optical disks. The new disks double the previous high water mark of 2.6GB on a dual-sided 5.25-inch disk.
These disks are formatted to hold 2,048 bytes in each disk sector, and that totals up 5.2GB per disk. As a result, HP`s SureStore Optical 80fx Jukebox packs 80GB into a rather compact and very reliable package. These RAID-array-sized jukeboxes house 16 slots and can be configured with either one or two optical drives. The model tested by IT/IS Labs had two drives. Naturally, the drives in the new jukebox are backward compatible with older 2.6GB, 1.3GB and 650MB disks.
The basic specifications for the HP SureStore Optical 80fx call for it to perform 600,000 disk swaps between failures. Sustained read I/O is pegged at 3.4MBps and sustained write I/O is pegged at 1.7MBps. With those sorts of bandwidth constraints, it should come as no surprise that the jukebox sports a standard 50-pin Centronics SCSI-1 interface. In real-life testing that included all of the overhead imposed by the Windows NT 4.0 operating system, IT/IS Labs found typical throughput for MO disks was significantly lower than the sustained high-end I/O limits.
We attached the HP SureStore Optical 80fx to a Dell 4200 PowerEdge Server via an Adaptec 2940 SCSI controller. The PE4200 server is a two-way SMP system powered by twin 300MHz Pentium II CPUs. We had an HP SureStore DLT tape library connected to this server via one of the server`s built-in Adaptec controllers. When we rebooted the server, Windows NT 4.0 recognized the two drives in the MO jukebox as two individual removable media drives and had no notion whatsoever of the jukebox--that won`t change until NT 5.0 ships.
Without software to leverage the jukebox, we were left to manually hit the buttons on the front panel of the jukebox in order to jockey disks in and out of the drives. Not a very satisfactory solution, but it did confirm the 12-second average switch time for a disk in the SureStore Optical 80fx. As for general I/O throughput, IT/IS labs measured write throughput on the order of 400KBps and read throughput on the order of 1.9MBps.
Next, we loaded Optical Server for Windows NT from Computer Associates on our server. This software makes the media in a rewritable optical jukebox appear as one or more large hard drives. CA-Optical Server transparently maps a group of MO media residing in the jukebox`s slots so that the group appears as a standard hard drive to the Windows NT operating system.
These virtual disks can be formatted as FAT or NTFS volumes. As a result, this virtual disk can be shared among many users and applications should be able to store and retrieve data using the virtual disk without any modification. All the standard Windows NT utilities, such as File Manager, Explorer, Disk Administrator, and Format, work with Optical Server.
When a user accesses data, Optical Server determines the media that is required, automatically mounts and unmounts media in drives, and fetches data when requested by an application. Optical Server enables flexible configuration of a jukebox into multiple volumes that can be dedicated to different users, projects, or administrative domains. In addition, the media that make up a single virtual disk can be exported from the jukebox and kept off-line. As a result, the effective capacity of the jukebox is much greater than the physical capacity as determined by the number of slots.
Another important configuration feature of CA-Optical Server is the ability to set up a hard-disk cache for a virtual MO volume. As the IT/IS Labs benchmarks demonstrated, this feature will be most beneficial when writing data. In our tests of CA-Optical Server, writes to the virtual disk were performed at hard-disk speeds and peaked at 3MBps as opposed to the 400KBps of the raw device.
The settings for this hard-disk cache for an MO disk are conceptually quite similar to the settings of a RAM cache for a hard disk. Watermarks correspond to cache-utilization levels and are used to trigger cache flushing. The defaults for these settings are at a glacial pace, however, when compared to those of a RAM cache. For example, the default cache flush period is 60 minutes.
When reads are performed on the virtual disk, the data is first staged on the hard-disk cache. As a result, when multiple users access the virtual disk and there is locality in the access pattern of data, then the data is accessed at hard-disk speeds.
Nonetheless, the time it takes to extract, flip, swap and insert new media can be prodigious when virtual disks span volumes. When we initialized the Windows NT disk administration utility on our server, it took a full six minutes to get through our jukebox. IT/IS Labs would highly recommend running CA-Optical Server on a dedicated system.
A Bit of the Blues
Isolating the current version (1.5) of CA-Optical Server is absolutely mandatory when it is used in conjunction with Adaptec`s 2940 PCI controller (AIC78XX.SYS) or Compaq`s embedded 710 or 810 SCSI controller (CPQFS32.SYS). In the installation release notes is a warning that a system connected to an MO jukebox using one of these controllers "may experience a system crash when Optical Server is running."
The release notes list the problem as residing with the drivers for these controllers. The real problem, however, is that having compatibility problems with these controllers is the equivalent of trying to sell automobile tires that have incompatibilities with the Ford Taurus.
All of the problems we encountered happened only when configuring virtual disks for the jukebox. When accessing data, either reading or writing files, CA-Optical Server was rock solid, When we were performing administrative tasks, however, the system was about as stable as a drunkard careening downhill on a skateboard. Whether creating a disk cache or formatting an MO platter, we were all too aware of the potential of being confronted with a "blue screen of death."
In addition, CA-Optical Server could not distinguish between the HP 80fx MO jukebox and the HP SureStore DLT library. Whenever we launched the CA-Optical Server administration tool, two jukeboxes always appeared in the configuration menu, with one being the DLT library.
Interestingly, none of these problems occurred when we ran our configuration with CA-ARCServe. We were able to configure the HP SureStore Optical 80fx and format disks without a problem. Moreover, ARCServe readily distinguished the SureStore Optical and the SureStore DLT as an optical jukebox and tape library respectively.
Throughput on reads was considerably higher than on writes and actually rivaled hard-disk performance. In the IT\IS Labs test, the MO drive was pegged at 1.9MBps.
Copying a file from a hard disk (2) to an MO virtual disk created by CA-Optical Server (1) shows the effects of a disk cache backing the MO virtual disk. Write performance to the disk cache is on the order of 3MBps. Note that while the cache fills, there is no activity on the MO disk (3).
This article is reprinted with permission from BackOffice magazine, a sister publication of InfoStor. For more information or to subscribe, visit www.backofficemag.com.