On 1 Feb 2001, at 22:42, Andrew Doran wrote:

> I'll borrow an Adaptec SCSI controller and do some benchmarking
> against a DPT one in the near future.

Hi,

that would be nice.

If someone is intersted in the tests I referred to:

I found that article again (a german computermagazine, iX from 
April 1999, page 112-118). The article can be found in german at : 
"http://www.ix.de/ix/artikel/1999/04/112/" (some Charts are missing 
in the printed original too)

The tested systems were:

Softwareraid:

Asus P2L97, PII 233MHz., 128MB Ram, AHA 2940UW, 5xIBM 
DCHS09U, Kernel 2.1.14, Raidtools 0.42

Hardwareraid:

Asus P2L97, PII 333MHz., 128MB Ram, 5xIBM DDRS 39130W, 
Kernel 2.1.x Raid controller ICP Vortex with 64MB Ram, 2 UW 
channels, i960 CPU. 

Both systems were tested with an adapted version of bonnie.

It was mentioned that the systems were not identical (two vendors), 
the softwareraid system had a slower CPU und slower disks, most 
parts of the filesystem were distributed as Raid 0 (/tmp and /usr) or 
Raid 5 (/home) over all five disks. The hardwareraid had one 
systemdisk the other four disks were used as Raid 5.

The findings:

With a filesize of up to 64MB the hardwareraid is much faster 
(because of its 64MB cache) then it becomes radically slower  until 
a 128MB filesize.

The softwareraid performs well up to 96MB, then performance 
drops until a filesize of 128MB.

At 128MB and higher filesizes the softwareraid is about 3 times as 
fast as the hardwareraid. This is explained with the distribution of 
the data to five drives, and possibly a slow Raid controller CPU.

The article states, that these tests must be judged very critical, as 
the systems were very differnet and ICP, the Raid controller 
manufacturer, had experienced very different results with some 
tests ( 29MB/s where this test had only shown 10MB/s).