It occurred to me that Thor Lancelot Simon wrote in gmane.os.netbsd.general:
> On Tue, Apr 12, 2005 at 06:58:50PM +0000, Martijn van Buul wrote:
>> 
>> How come sd0d is so slow sometimes?
>
> When you dd to and from the block device node, every I/O forces allocation
> in the metadata cache.  Why would you expect that to be a good thing?

I expected it to be slower - but not *THAT MUCH* slower. 

When I run this on my ATA disk, I get 14627040 bytes/sec for /dev/wd0d. That's
about as fast as the raw sd0 drive - which is a slower drive, over a slower
interface. I would think that the impact of having to allocate in the metadata
cache would be more or less the same for both drives.

Another thing is that newfs (without -O 2) is painstakingly slow. I aborted
it after 10 minutes. Creating a ufs2 partition was much faster.