On Thu, 2007-11-29 at 21:39 +0100, moi wrote: > RAID5 (if it really is one) ALWAYS has one drive´s capacity as spare... > the spare blocks are just distributed on the disks, thus avoiding the > bottleneck of a single spare drive (these would be raid levels 3 and 4). > > what you meant was RAID6/ADG, a semi-proprietary stuff rather found on > hardware controllers, e.g. hp smartarrays. these do calculate a parity > for each n blocks, and for "n blocks+parity" generate a second parity > block. All these blocks are distributed evenly on all drives in the array. > > The thing with ADG is the rebuild time - for example the RAIDs at work > have about 20 drives each (300 gig); the rebuild time on those is about > 1 gb per hour minimum (when there is heavy activity on the raid set). > that would mean 300 hours without any protection (when using raid5) ! > instead, with raid6/adg there still is one parity left. > bad thing, though, is the raid controller has to calculate a lot of > parities. furthermore, the cost is rather high with 2 disks´ worth of > parity. Most of the time, such setups use RAID10 (mirror and stripe), > which uses much cheaper controllers and offers more performance. > > sorry for off-topic :) Two remarks: A. Modern RAID5 (be that software and/or hardware controller) build far faster then 1GB/h (291KB/s!??!?!). I timed my own sever (6 250GB drives in software RAID5) at ~12MBps (42GB/h) load and ~90MB/s (324GB/h) idle. B. The Linux kernel has built in software RAID6 support; while slower then the RAID5 implementation, the performance hit is noticeable but not devastating and given the added price (1 250/320/etc GB SATA drive) RAID6 is indeed a fair option if you require two-failed-disk support. - Gilboa
