[ale] raid suggestions

[Pat Regan]

On Thu, 9 Feb 2012 20:37:03 +0000
"Lightner, Jeff" <JLightner at water.com> wrote:

> Hardware based (for true RAID) because it doesn't require CPU cycles.

I'm not sure what your definition of "true RAID" is.  As for software
RAID requiring more CPU cycles, though, the number of cycles required
is quite negligible on anything approaching a modern processor.

Here's how fast the kernel says my laptop can crunch the numbers for
RAID 6:

[428688.926111] raid6: int64x1    793 MB/s
[428688.993945] raid6: int64x2   1052 MB/s
[428689.061905] raid6: int64x4    713 MB/s
[428689.129830] raid6: int64x8    695 MB/s
[428689.197810] raid6: sse2x1    2263 MB/s
[428689.265749] raid6: sse2x2    2649 MB/s
[428689.333668] raid6: sse2x4    2941 MB/s
[428689.333675] raid6: using algorithm sse2x4 (2941 MB/s)

That is just shy of 3 GB per second and I'm pretty sure that these
calculations are single threaded.

CPU use isn't the performance argument to make against software raid,
though.  Bus bandwidth is usually a bigger factor.  Mirroring requires
almost zero cpu but it requires double the bandwidth.  This was a big
deal when your PCI bus was limited to only about 130 MB/s.

Today that isn't usually worth worrying about.  Even a single PCIe lane
has enough bandwidth for a pretty large number of drives...

Pat