Pathological RAIDframe rebuild times

[Stephen Borrill <netbsd%precedence.co.uk@localhost>]

I've a couple of HP Microservers, both with 4x WD Caviar Green 2TB drives. 
Operating system (netbsd-5) is on USB (same image used on both). Drive 
write caches are enabled in the BIOS and also double-checked with dkctl.

I'm attempting to create a RAID5 array, stripe is 32 and partitions start 
at 0 sectors. I've tried both 3 and 4 disk arrays (understanding the 
preference for 3 disks ones as discussed in earlier threads).

All drives have had a full dd of /dev/zero over them before starting (hit 
lots of bad sectors out of the box on this model of drive).

On one server, the build took around 12 hours (which is similar to seen in 
the past). On the second, it estimated around 120. After a bit of playing 
around (nothing specific, just starting and stopping, changing from 4 to 
3 disks, double-checking BIOS settings), a new build said 12 hours. But 
after about 10% it ramped up to 120hours again.

tps on the disks is around 200. Transfer rate (according to systat iostat) 
is around 3.5 MBps. Interrupt rates are around 650/s.

The drives are doing little: a simultaneous dd read gives around 65MBps 
(compared to double that for the spare drive).

--
Stephen