Capacity · tolerance · rebuild risk
RAID calculator. Usable capacity, what it tolerates, and an honest estimate of the rebuild.
Pick the level, the number of drives and their size, and the drives' rated unrecoverable read error figure. Back come the usable capacity, how many drives the set tolerates, and an estimate of the chance a rebuild meets an unrecoverable read error on the survivors, with the caveats that estimate needs. It is a planning tool and a warning, not a diagnosis: if a rebuild has already failed, the rebuild-failed page is the one to read, and the first look is free.
Rather talk it through? An engineer answers the bench line
0800 6890668
Usable capacity, tolerance, and the rebuild estimate.
What each level tolerates, in words →The level
The drives' rated unrecoverable read error figure, one error per this many bits read
What the numbers mean, and what they do not.
Usable capacity is the drives' capacity less what the level spends on redundancy: nothing on RAID 0, half on RAID 1 and 10, one drive on RAID 5 and RAID-Z1, two on RAID 6 and RAID-Z2, three on RAID-Z3, and one or two per span on RAID 50 and 60. Decimal terabytes are used throughout, as on the drive's label; a file system will show a little less.
Tolerance is how many drives can fail before the data does, in the worst case: one per mirror pair on RAID 10 is shown as one, because the second failure can land in the same pair.
The rebuild estimate is the naive reading of the drives' rated unrecoverable read error figure. A rebuild reads every sector of every surviving member; the data read is the survivors' capacity, in bits; the expected number of errors is that divided by the rating, and the chance of at least one is one minus e to the power of minus that number. A drive rated at one error per 10^14 bits is rated to meet one in about 12.5TB read, so a rebuild of four 8TB drives, reading 24TB, expects nearly two on paper. The rating is a specification ceiling, not a measured rate: if it were a real rate, most drives could not complete a long sequential read, and a 2013 paper on reliability models put UREs between 10^-14 and 10^-15 per bit while noting that the arithmetic gave an eight-plus-two RAID 6 of 1TB drives only about a 53 per cent chance of reading its survivors clean after two failures. Treat the estimate as an upper bound and a reason: a reason rebuilds fail late, a reason to image the survivors before rebuilding a set with a weak one, and a reason for RAID 6 or RAID-Z2 on large drives. It is not a prediction for any one array.
What the calculator cannot see is the thing that decides most rebuilds: whether a survivor already has pending sectors. SMART attribute 197 on a SATA drive, or the grown defect list on a SAS drive, is a count of sectors the drive cannot currently give, and a rebuild asks for every one of them. A survivor with pending sectors fails the rebuild at the first of them whatever the URE rating says. The rebuild-failed page is what happens next, and the drive site's predictive-failure page is what the numbers mean.
The questions that come up first.
Is RAID 5 dead?
For large drives, it is the wrong default. The URE arithmetic is a ceiling, but the direction is right: the more data a rebuild reads, the likelier it is to meet a sector a survivor cannot give. RAID 6 or RAID-Z2 costs one more drive and tolerates the error.
Why does the estimate say more than 100 per cent expected errors?
Because the expected number of errors can exceed one; the probability of at least one is capped below 100 per cent by the formula. Both are shown.
Does the estimate apply to a mirror?
A mirror's rebuild reads one drive, so the data read is one drive's capacity. The estimate is shown for that.
My rebuild already failed. What now?
Power down, keep the drive that failed first, and read the rebuild-failed page. The set is usually recoverable from images.
What does recovery cost if it goes wrong?
A set of two to four members is £500 + VAT upwards after the free look, fixed in writing; larger sets from £1,250 + VAT.
If the rebuild has already failed, the numbers are behind you.
Power down, keep every member, and send the form. The first look is free, and the set is usually recoverable from images.