RAID Scaling Charts, Part 1

Test Setup

Swipe to scroll horizontally
System Hardware
ProcessorAMD Opteron 875 (Egypt 90 nm, 2.2 GHz, 2x2 MB L2 Cache)
MotherboardIwill DK88 Dual Opteron BoardChipset: Nvidia nForce Professional 2200, BIOS: 1.4
RAM2x 1024 MB DDR-400 (CL 3.0-3-3-8)Micron 18VDDF12872G-40BD3
Graphics CardATI RagelXL
System Hard Drive120 GB 7,200 RPM, 8 MB Cache, SATA/300Samsung HD120IJ
Benchmark Hard Drives8x 320 GB 7,200 RPM, 8 MB Cache, SATA/300Samsung HD321KJ
RAID ControllerAreca ARC-1220 8-Port PCIe RAID6-ControllerFirmware V1.43 2007-4-17
ATI GraphicWindows integrated
Nvidia nForce ProfessionalVersion: 6.70
OSWindows Server 2003 Enterprise Edition, Build 3790 SP2
Benchmark Software
Performancec’t h2benchw 3.6
I/O PerformanceIOMeter 2003.05.10Fileserver-BenchmarkWebserver-BenchmarkDatabase-BenchmarkWorkstation-Benchmark

All tests were performed at controller defaults.


All tests were run with two to eight Samsung HD321KJ, but using only 80% of the capacity; we did not use the full capacity of 320 GB per drive, in order to keep the test runs short. This reduced the test run times and made sure that the drives operated at maximum performance, because they never reached the inner sectors of the rotating media, which deliver the lowest transfer rates.

It took four to six hours to complete one benchmark session for one RAID configuration, including throughput, access time and I/O benchmark runs. Initialization was required for all RAID arrays except the RAID 0 setups; this took an additional 20 minutes.

We also made a nice discovery with regard to drive power supply: we first attached multiple drives to a single power rail of the power supply used. This seemed to work well, as the spinup typically draws enough power to cause trouble and we didn’t encounter problems. However, single drives dropped out of the RAID array during the benchmarks once in a while. We blamed the drive when it first happened, and pointed at the RAID controller after experiencing the problem a second time. Then we split the power supply to three different rails and voilá - the problems disappeared.