WD has been sampling SMR - Shingled Magnetic Recording - disk drives to several cloud vendors including Facebook, Google, Microsoft and Amazon. In September Seagate announced sales of a million SMR drives - I didn't know they'd sold any - and is planning a 5TB SMR drive for next year. HGST - an independent unit of WD - likely has a similar program under way.
What is SMR?
Disk drive tracks are separated by a gap. In SMR drives, this gap is removed to roughly double density, says a paper from CMU's Parallel Data Lab.
Write heads lay down a wider track than read heads need. In an SMR disk the write tracks are overlapped, leaving narrow tracks that are fine for the read head, but which can't be overwritten without destroying the data on adjacent tracks.
In practice, the tracks on an SMR disk are laid down in bands of tracks. The band enables a partial rewrite of the disk by way of an expensive read-modify-write cycle like that used in RAID 5 arrays.
Facebook is testing Western Digital SMR drives in its new 65,000 square foot cold storage data center in Prineville, Oregon. FB found that 8% of photos got 82% of the accesses, which means over 90% of the photos share the remaining 18% of accesses.
FB knew that few cold photos would ever be accessed, but they still wanted them available just in case - which ruled out tape. Storing them on SMR disks maximizes capacity at little extra cost.
FB has published a spec - Cold Storage Hardware (pdf) - that illuminates the limits of SMR drives. The big issue: SMR drives are very sensitive to vibration - only 2 can run concurrently in a 30 drive shelf - which is probably why there's no marketing aimed at consumer or enterprise prospects.
Another issue is managing the rewrite process. Should the intelligence be in the drive or on the host? WD is taking the "dumb drive" approach - let Facebook's many software engineers worry about it - rather than handling rewrites within the drive, which I expect is the Seagate approach.
The cloud vendors probably favor the dumb drive approach because then the band size can be larger - using server resources - that a disk controller could handle, giving them more useful capacity at the cost of some CPU cycles. But perhaps Seagate has some more tricks up its sleeve.
The Storage Bits take
I've been agitating for specialized archive disks for years and it looks like we're getting close. Low duty cycle drives need to be engineered differently - for example, lubricants - and SMR density is helpful too.
The vibration problem is serious, because vibration has been a disk problem for decades and is only managed, not solved. It may not be possible to design a reliable consumer SMR system at an affordable price.
But the advent of a- and the promise of a reliable archive Blu-ray - means that perhaps consumers won't want tricky 12TB archive disks. Yet our digital civilization needs reliable digital archives - and competing SMR drives from WD and Seagate are a step in the right direction.
Comments welcome, of course. Good to see that drive competition isn't dead.