We invited each vendor to supply a disk array subsystem with five drives that have a total capacity of 10 to 12 GB, configured as a RAID 5 array. Although it wasn't required of them, some vendors also supplied a sixth drive to act as a hot spare. We specified that the subsystem's interface to the host be SCSI-2 Fast/Wide. Of the 16 products tested, 10 had RAID controllers built into their enclosures; eight of these products supported a SCSI-2 Fast/Wide single-ended termination and the other two supported differential termination. To connect the single-ended subsystems, we installed an Adaptec AHA-2940W PCI-to-Wide SCSI adapter in our test file server. To connect the differential products, we installed an NCR 8251D PCI SCSI adapter. The remaining six arrays shipped with their own RAID controller boards, which doubled as host adapters for these
products.
We used a Dell PowerEdge SP590-2 system as our file server. The PowerEdge is a Pentium 90-MHz-based EISA server with two PCI (peripheral component interconnect) slots. Microsoft Windows NT 3.5 Workstation was installed on the boot drive of the Dell. We evaluated each product's performance, usability, and features, and the test results were weighted as follows: 50 percent, 25 percent, and 25 percent, respectively.
PERFORMANCE
We connected each disk array we were testing to the file server using the appropriate host adapter. We then formatted the array under NT as a single drive using the NTFS (NT File System) format. We ran a suite of performance tests under NT with the array's write-back cache off and then on (if both states were supported and could be toggled by the end user).
The performance suite simulates tasks that a disk array subsystem would perform in a real-world environment. Random and sequential reads and writes of 4-, 16-, and 64-KB blocks w
ere performed at different locations on the array in a single-thread and a multithread environment. Except for the tests that read or wrote over the disk array, we set the number of blocks per segment so that the total size of the region under test was 128 MB.
PERFORMANCE SCORING
We recorded test results as the average, minimum, and maximum time (in seconds) required to complete each test. The average and maximum times gave performance scores; minimum times were for reference only. A product's score is relative to how it performed compared to the other products. Each product's Best for Database Servers score is a weighted average of the single-thread and multithread "average" recorded times. The Best for Audio/Video score is a weighted average of the single-thread and multithread "maximum" recorded times. The Best Overall score is an average of the database and audio/video scores. We used a product's "cache-on" times if the product was supplied with a battery-secured write cach
e; otherwise, we used the "cache-off" times.
FEATURES
We evaluated each product on its cost per MB of storage, warranty length and coverage, redundant and hot-swappable components, as well as alarms, security features, and maximum storage capacity.
USABILITY
We evaluated each product's ease of setup and configuration and the completeness and clarity of the user's manuals. We simulated a single-drive failure, verified that the file server could continue to operate normally, and evaluated the ease of performing a rebuild of the array.
Contributors
Michele Guy, Project Manager/NSTL,
has been testing hardware and software products for NSTL for the past four years.
Kathleen Bishop, R&D/NSTL,
has eight years of R&D experience in the computer industry.
Bruce Levy, Ph.D., Manager, R&D/NSTL.
The Lab
How We Tested
