The disk industry's fight with flash storage is hampered by disk areal density flatlining the past 5 years. Instead of disk capacity doubling every 18-24 months as it did in the 2000s, it's creeping up at approximately 20 percent per year.
The long-term answer is two-fold: patterned media; and heat-assisted magnetic recording, or HAMR. But scaling up lab demos of patterned media to economically producing several billion platters a year is expensive and years away. HAMR has similar challenges.
Thus the interest in shingled magnetic recording (SMR).
SMR for the curious
Disk heads write a wide track, but only need a narrow track for reading. Thus shingled recording: write the wide track; then overwrite most of it on the next pass, leaving a narrow track for reading.
In theory this could double today's disk capacities. But there is a hard problem: updating.
Like flash memory, updates require reading the existing data, integrating the new data, and then writing back to the disk. Slow and costly, the same write amplification problem found in flash.
Great for archive disks — where SMR is used today — but not so great for frequently updated data. How to fix?
Novel SMR address mapping
In a recent paper presented at the Usenix Hot Storage 2014 conference Weiping He and Prof. David Hung Chang Du from the University of Minnesota's Center for Research in Intelligent Storage presented their take on the problem.
Novel Address Mappings for Shingled Write Disks offers a partial solution to the problem. Instead of writing each track in order — track 1, 2, 3, 4 — write them out of order: 4, 1, 2, 3.
When the disk — or the band that contains these tracks - is less than half full, tracks 4 and 1 can be freely updated as there is no shingling. Until 75 percent of the disk is used, tracks 2 and 4 can be freely updated.
The paper explores various permutations on this theme and concludes
"By appropriately changing the order of space allocation, the new mapping schemes can improve the write amplification overhead significantly. [...] [N]ew mapping schemes provide comparable performance to that of regular HDDs when SWD space usage is less than 75 percent.
The Storage Bits take
Double the nominal density of hard drives with SMR and use 75 percent of that means a 50 percent boost in capacity. Not bad for firmware tinkering.
Integrating this concept with a non-volatile RAM buffer would enable even higher capacity utilization of the SMR drive.
We haven't seen the last of the hard drive.
Comments welcome, of course. How full are your drives on average?