Most testing was done using Intel's IOMeter benchmark software. Although discontinued by Intel, it remains by far the best way to measure raw I/O performance, and when properly set up correlates well with many typical server-side applications. (In contrast, most desktop apps exibit disk operations that are highly localized, and as such are more accurately modeled by benchmarks like Intel's IPEAK.) We created profiles to showcase the so-called "four corners" of low-level disk performance -- sequential read, sequential write, random read, and random write -- along with several profiles targetted at the likely uses of a massive storage subsystem. In detail:

• Database -- a profile designed by Intel to mimic the I/Os handled by a typical database.
  • 2KB reads, 67% random
• Video Capture -- an application that heavily stresses write speed, possibly on a fragmented drive.
  • 64KB writes, 5% random
• Video Server -- an application becoming more common on SOHO networks, stressing several drive parameters. Profile obtained by monitoring actual video streaming via Sysinternal's [http://www.sysinternals.com/ntw2k/source/filemon.shtml∞ FileMon] tool.
  • 90% 64KB blocks, 67% reads, 10% random
  • 5% 16KB blocks, 67% reads, 50% random
  • 5% 4KB blocks, 67% reads, 100% random
• File Server -- another profile designed by Intel, this time patterned after a general fileserver.
  • 10% 512b blocks, 80% reads, 100% random
  • 5% 1KB blocks, 80% reads, 100% random
  • 60% 2KB blocks, 80% reads, 100% random
  • 2% 8KB blocks, 80% reads, 100% random
  • 4% 16KB blocks, 80% reads, 100% random
  • 4% 32KB blocks, 80% reads, 100% random
  • 10% 64KB blocks, 80% reads, 100% random
• Web Server -- serve up your GigaCollection of digital toys to the world, or just keep a weblog. Profile designed after a well-known statistical model of the Nagano Olympics website by IBM [1].
  • 80% 4KB blocks, 100% reads, 85% random
  • 10% 4KB blocks, 90% reads, 100% random
  • 10% 64KB blocks, 100% reads, 10% random

Base controller performance was assessed used only the "four corners" tests. Cable integrity was tested using the same IOMeter benchmarks as complete arrays; also included was a bulk copy of binary data across both channels, repeated 10X, and then examined for bit-for-bit accuracy.

TheTests