Re: Is this enough for us to have triple-parity RAID?
|[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
On 18/04/12 20:22, Piergiorgio Sartor wrote:
Hi David, On Tue, Apr 17, 2012 at 10:18:55PM +0200, David Brown wrote: [...]For quad parity, we can try g3 = 8 as the obvious next choice in the pattern. Unfortunately, we start hitting conflicts. To recoveryou should not use 8, because this is not a generator of GF(256) with polynomial 285, the standard for the RAID-5/6 setup. This means than 8^k does not cover the complete field for k in [0 254], thus having cycles and, consequently, creating conflicts. Some generators could be: 2, 4, 6, 9 13, 14, 16... but not 32 nor 64. I know that powers of two are nice, but if you want to have generic RAID, you must use other values.
I know that 8 is not a generator, and therefore you cannot expect to get a full set of (256 - noOfParities) disks. But picking another generator (such as 16) is not enough to guarantee you the full range - it is a requirement, but not sufficient. The generators need to be independent of each other, in the sense that all the simultaneous equations for all the combinations of failed disks need to be soluble.
It turns out that if you pick 16 as the forth parity generator here (1, 2, 4, 16), then you can only have 5 data disks. In fact, there are no other values for g3 that give significantly more than 21 data disks in combination with (1, 2, 4), whether or not they happen to be a generator for all of GF(2⁸).
The log/exp tables, are, of course, always valid. BTW, the GF(256) with polynomial 285 has exactly 128 generators, so it would be possible to have up to 129 parity disk (1 is not a generator), for, I guess, a max of 256 disks (or maybe 255?). Hope this helps, bye,
When I started out with this, I thought it was as simple as you are suggesting. But it is not - picking a set of generators for GF(2⁸) is not enough. You have to check that all the solution matrices are invertible for all combinations of failed disks.
In fact, it is a little surprising that (1, 2, 4) works so well for triple parity. I don't know whether it is just luck, or genius in Peter Anvin's choice of the multiply operation.
mvh., David -- To unsubscribe from this list: send the line "unsubscribe linux-raid" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html