RAID5 on a new microserver?

[abaxas]

This myth has to die. CPU power hasn't been relevant to software RAID performance for a decade. The CPU in the N40L can calculate parity at ca. 4GB/sec (in Linux). It's not a bottleneck.

For reference, my N40L will do a "zpool scrub" (roughly equivalent to a RAID rebuild, though much more CPU intensive) at around 100MB/sec per drive. Which is about all the old Spinpoint F1s have to give.

Not strictly true.

Raid rebuild time are drastically increased if the device is 'in use' during rebuild. Which by definition is risk to data. It's the seeks that kill you.

[b0redom]

Are you sure? I had a very large (I think 8x 1.5TB drive) RAID-5 array on a Perc6i hardware RAID controller. Building the array took hours. It was much quicker using ZFS, but when I rebuilt my new ZFS array following a disk failure (5x2TB RAID-Z), it was still (lots of) hours.

[drsmithy]

Not strictly true.

Raid rebuild time are drastically increased if the device is 'in use' during rebuild. Which by definition is risk to data. It's the seeks that kill you.

Yes, but that has (next to) nothing to do with the CPU performance, which was the topic of the comment I replied to.

Unless the CPU utilisation on the Microserver is sustained at upwards of 80%, RAID rebuild/sync performance shouldn't be impacted.

[drsmithy]

Are you sure? I had a very large (I think 8x 1.5TB drive) RAID-5 array on a Perc6i hardware RAID controller. Building the array took hours. It was much quicker using ZFS, but when I rebuilt my new ZFS array following a disk failure (5x2TB RAID-Z), it was still (lots of) hours.

Very sure. Your minimum rebuild time is, basically, dictated by the speed of a single drive [0], which will be somewhere in the ballpark of 80-120MB/sec. Since the CPU can checksum at ca. 4GB/sec, it's obviously not going to be the bottleneck in the process.

ZFS rebuilds faster because it only rebuilds the portion of the disks that's in use, rather than the whole thing (some traditional RAID setups also do this, but not the PERC6, so far as I know).

[0] Used to be bus speed also played a part, but these days with PCIe it's practically irrelevant (unless you're doing something silly like hanging half a dozen drives off a single PCIe x1 card).