Compaq Smart Array 3200 Controller as a SCSI Controller
paulkoning at comcast.net
Fri Jul 17 15:39:12 CDT 2020
> On Jul 17, 2020, at 4:19 PM, Grant Taylor via cctalk <cctalk at classiccmp.org> wrote:
> On 7/17/20 12:58 PM, Paul Koning via cctalk wrote:
>> What is "three drive parity"?
> A poor choice of words in a complex topic.
> How about "three drives worth of redundancy" Meaning that your data will still be accessible if three drives fail.
> ZFS has three versions of ZRAID or RAID-Z.
> - RAID-Z1 is analogous to RAID-5.
> - RAID-Z2 is analogous to RAID-6.
> - RAID-Z3 is analogous to ???
> I'm not aware of any official definition of a mirror of more than two drives. I've heard of "3-way" / "4-way" / "n-way" mirrors.
> I think that the industry has settled on RAID-10 / RAID-01 and possibly RAID-11 / maybe even RAID-00. But that isn't a standard to me.
> Further, I see disagreements of what is the strip and what is the mirror in RAID-10 / RAID-01.
>> "Parity" is the description of RAID-5, and 3 drive RAID-5 is certainly perfectly standard. RAID-1 is not parity, it's mirroring.
> If you think of it as "redundant drives" or "number of drives that can fail without destroying data", then yes, RAID-1 does have a numerical value quite similar to RAID-3 / RAID-5 / RAID-6 / RAID-Z* / etc.. Though nomenclature becomes highly problematic.
Yes. Marketeers like to use the phrase "erasure coding" as a magic token to indicate great wonderfulness beyond RAID. Actually, it's a math term that describes the standard RAID systems plus generalizations. I don't know if there is a clear terminology for the general thing. I tend to call them "k of n" codes, k data bits, n message bits. Or in the RAID case, k drives worth of data, n total drives. That makes RAID-1 a 1 of 2 erasure code, the thing we started talking about 1 of 3 erasure code, RAID-5 is n of n+1 erasure code, and RAID-6 is n of n+2 erasure code. So in the k of n notation, n-k drives can fail before you lose your data.
Actually, more realistic with today's drive sizes: n-k-1 drives can fail and one drive can have a hard read error during reconstruction before you lose data. For sufficiently large drives, the probability of a hard read error somewhere is nearly 1, which is why RAID-5 is only a good idea for small disk drives.
People have build storage systems with other erasure codes, for example n of n+3. Not because it's really necessary but as a research experiment; the "Self-star" system at CMU comes to mind.
>> Is the question about triple mirroring, i.e., 3 drives all having the same data on them?
> I was stating that I'm not aware of an official RAID level designation for ZFS's RAID-Z3.
>> That's pretty rare though not unheard of, I've never seen a RAID-x designation for that.
> I've known more than a few people to use n-way mirrors (~RAID-1). Though I think I've only seen it in software.
>> For high availability, RAID-6 is much more economical (and at this point the standard choice); triple mirroring is of that class, with the difference that it performs better for random short writes.
> Are you comparing RAID-6 to triple (3-way) mirroring? Or something else?
> I think that things get really weird and deep in minutia when you start comparing a 3-way mirror to a 3 drive RAID-6. Same number of drives (3), and same capacity (1 drive worth), and same fault tolerance (2 drive failures).
Yes. The difference is that the 3 drives in a 3-drive RAID-6 system aren't copies, so the write and reconstruction logic is more complex.
But actually I meant RAID-6 generally. As I mentioned above, in RAID-5 if you lose a drive, you have no redundancy, which means that if you encounter a hard read error during reconstruction you've lost that stripe. And for today's large drive chances are that will happen. With RAID-6, if one drive fails you can still recover from hard read errors.
If you run the statistical analysis with typical drive MTBF and hard read error rates, you'll find that the read error during reconstruction is the major contributor to data loss, not the "too many drives failed at the same time" scenario. At least not if the drive chassis is built correctly and the power source and other environmental parameters are reasonably clean.
More information about the cctalk