I think ECC memory is a good idea if you can get it; turning memory errors from "random data corruption and crashes" into "harmless fmd/syslog messages" is pretty nice, and last time I looked, would add about £5 to a 4GB DDR2 system.
Of course the problem then is that Intel don't support ECC on anything but Xeons; AMD are rather better, all you need is a supporting motherboard (which includes some cheap and cheerful integrated ones), any AMD CPU in the past few years will do.
Now, if only Intel would get off their arses and release the W3xxx Xeons so I can build an ECC protected i7 system. 12GB without ECC is getting a bit ridiculous.
I'd also go for a pair of dynamically striped mirrors instead of RAID-Z2; modern disks can already pretty much saturate GigE individually, so I'd rather go for increased overall IOPS rather than increased sequential throughput at the cost of making every drive take part in each IO.
Yup, pretty much; if you have more than one vdev in a pool (mirror, RAIDZ, whatever), ZFS will stripe over them, varying what goes where depending on how fast it thinks they are.
I've heard reports of ZFS doing things like putting more data on one vdev because the disks underlying it had 16MB of cache and those in the other vdev only had 8MB; bit fancier than what you normally associate with RAID10.
Of course the problem then is that Intel don't support ECC on anything but Xeons; AMD are rather better, all you need is a supporting motherboard (which includes some cheap and cheerful integrated ones), any AMD CPU in the past few years will do.
Now, if only Intel would get off their arses and release the W3xxx Xeons so I can build an ECC protected i7 system. 12GB without ECC is getting a bit ridiculous.
I'd also go for a pair of dynamically striped mirrors instead of RAID-Z2; modern disks can already pretty much saturate GigE individually, so I'd rather go for increased overall IOPS rather than increased sequential throughput at the cost of making every drive take part in each IO.