To Worry Or Not To Worry

Or, Much Ado About Random Write Endurance

Intel's 320-series 300GB SSD has a quoted 4KB random write endurance - that is, the minimum total volume of data you can write to it in individual 4KB randomly located blocks before it begins to fail - of 30TB.

30TB may sound a lot to you.  The primary MySQL server at my day job does 2.5TB of writes per day (and it's only one of several database servers).  MySQL writes tend to be random-ish, so you might at first glance expect the abovementioned drive under those conditions to burn out in 12 days.  For that reason (and the fact that the database is rather larger than 300GB), we don't use a 320-series SSD; we use a RAID-50 array of 20 enterprise drives each with about 60x the quoted write endurance.  Based on the quoted numbers and measured load, we should be good for at least 10 years.

The question is, though, what is the real-world longevity of SSDs under heavy random write conditions?  I've been very conservative about SSD deployment - for mee.nu I've used the more expensive enterprise SLC drives as well (and RAID-5 at that) even though our write activity is a couple of orders of magnitude lower.  The only MLC drives I've deployed in a production environment have been in applications where reads are random and writes are sequential - some of the Cassandra and Xapian databases at my day job fit this description.

However, this paper, presented at last year's Hot Storage conference, suggests that things might not be nearly so bad.  The authors examine a model of flash cell burnout, and note that if cells are given time to rest between write/erase cycles, their endurance can be expected to increase significantly.

How significantly?  Let's take our 300GB SSD and hit it with 2.5TB of data a day.  Let's assume a worst-case scenario on two  aspects - all of that is individual 4KB random writes, and there's no write-combining done by the OS or RAID controller.  Let's assume a best-case scenario on the other aspects - write multiplication is 1.0 (that is, no blocks need to be moved to allow for the updates) and wear-levelling is perfect across the drive (all blocks are updated evenly).  (All of these assumptions are completely implausible, but the idea is that they'll kind of balance out until I can get more precise data.)

That means that every block on the drive is updated every three hours.  A litle less than three hours, but near enough.  That paper suggests that with a 10,000 second - a little less than three hours - recovery period between write/erase cycles, write endurance of MLC cells can be expected to be 90 times the worst-case situation the manufacturers cite.

That is, rather than two weeks, the drive would last for three years.  And then drop dead all at once given our rather unreasonable scenario.

Which is a completely different picture from what the manufacturer's worst-case numbers might suggest.  And with a RAID controller with battery-backed write-back cache, the number of writes that actually hit the SSD can be significantly less.

The problem is, this is a simulation.  It's a very careful simultion based on the known physical properties of the semiconductor materials used in flash fabrication, but it's still a simulation.  I'm hoping I can get a couple of SSDs solely for the purpose of killing them, because I haven't seen anyone else publish good data on that.

The reason all this matters is that where a 300GB Intel MLC drive costst $600, 300GB of Intel SLC enterprise SSD storage comes to five drives totalling $4000.  The point may become somewhat moot when Intel's 710 MLC-HET drives launch.  The HET, I would guess, stands for something like high-endurance technology; these drives are based on cheaper MLC flash but optimised for reliability rather than capacity.  They will likely (based on reports in the trade press) cost twice as much as the regular MLC drives, but offer 20 to 40 times the endurance - nearly as good as SLC.  If the price and endurance turn out that way, then there will be 3x less reason to risk your data on a statistical model and a consumer drive.

Another thing: Intel's 320 series (unlike the earlier M-series) implement internal full-chip parity in the spare area, so even if one of the flash chips dies completely, the drive will continue operation unaffected.

Posted by: Pixy Misa at 03:56 AM | Comments (7) | Add Comment | Trackbacks (Suck)
Post contains 746 words, total size 5 kb.