One of the more surprising facts revealed in this article is that modern HDDs don't go through the traditional CHS order with increasing LBAs, and the number of sectors per track can vary even between two sides of the same platter at the same radial distance. I suspect a lot of this information is kept secret by HDD manufacturers, and known only to them and those working in data recovery (if a drive has one dead head, or media damage on one surface, traditional CHS order vs. the others would show very different results in which portions remain readable.)
It would be easy for competition to reverse engineer the physical layout.
I think it's "secret" because nobody really cares.
Do you care for example how the CPU maps the linear memory space across physical RAM sticks and chips? I recently discovered that there multiple alternatives and BIOS settings which give you some control over this.
You can't cheat physics, so at today's CPU speeds, spinning disk HDDs more closely resemble tape drives than not. You have to pay a significant performance price when moving the read head across the disk surface, moreso when you have to wait an entire rotation for the correct sector to fly under the read head.
If I were designing drives, I would be considering the platter and all the mechanicals to be noisy, bendy, sticky, shaky, fade-y, and all the things you wished your mechanical parts weren't. Every part of the disk has different amounts of noise, shakiness, etc, since the manufacturing process isn't consistent.
Yet you want to store as much good data on the whole disk as possible, and need to be able to retrieve smallish chunks of the data back in the typical case quickish.
In turn, that constrains how you group data, and how you allocate ECC bits. As a seller of disks, it's probably how you decide if you should stick a 3TB, 4TB or 5TB sticker on the box too.
The end result will be an algorithm that, based on factory calibration and recent measurements decides how to map data to physical areas of the disk. One head giving a weaker signal lately? Well we're gonna assign fewer data bits there.
And when reading data back there are also decisions to be made. For example, when we read a bunch of data but the ECC recovery fails due to too much noise, should we re-read the same data and combine the reads to get more signal and less noise? Should we shake the head a little and see if we can re-read and get better alignment next time? Should we go and read a few tens of megabytes of other data and see if we can recover the data from a higher hierarchical level ECC/erasure coding scheme?
I wonder if this approach still works with firmware tweaked drives. Anyone remember the "google" spec 320G Seagate drives from circa 2006? The ones with the pink epoxy"? IIRC they benchmarked with flat throughput across the whole platter. Rumours were they were custom spec drive that somehow made it in to the retail channel.
When I read it years ago the most surprising to me was density of 40MB IDE drive. ST-157A sold all the way to 1992 was barely more sophisticated than a DD floppy disk. ~10000 bpi while DD floppy is ~9000. Heads are narrower and there is more tracks per surface, 650 vs 135 TPI, and rigid magnetic surface lets it spin faster, thats it.
One of my cow-orkers did some work back around 2003 based on the finding reported in this paper:
"Trail : A Fast Synchronous Write Disk Subsystem Using Track-Based Logging" by Tzi-cker Chiueh and Lan Huang
...but I'm having a hard time finding a non-paywalled version now since many of the URLs in the search results have suffered bit-rot. Those findings dovetail with those mentioned here in that he used the true/discovered disk geometry map to allow him to stream data onto the disk very fast by writing to the specific sectors that were known to be arriving under the heads RIGHT NOW, which caused the data to appear out-of-order when accessed later via sequential LBA block numbers but could be reordered using that same map. (If somebody finds a non-paywalled version, please share...)
14 comments
[ 5.1 ms ] story [ 43.7 ms ] threadWouldn't that cause all sorts of issues with SMR drives?
In addition, the heads have multiple levels of independent actuation beyond the shared arm itself[1][2].
So as long as the tracks themselves are fairly consistent, I don't see why the number of sectors per track has to be equal.
[1]: https://pcper.com/2009/09/wd-unleashes-nested-actuator-and-a...
[2]: https://blog.westerndigital.com/hdd-magic-20tb-18tb/
I think it's "secret" because nobody really cares.
Do you care for example how the CPU maps the linear memory space across physical RAM sticks and chips? I recently discovered that there multiple alternatives and BIOS settings which give you some control over this.
You can't cheat physics, so at today's CPU speeds, spinning disk HDDs more closely resemble tape drives than not. You have to pay a significant performance price when moving the read head across the disk surface, moreso when you have to wait an entire rotation for the correct sector to fly under the read head.
Yet you want to store as much good data on the whole disk as possible, and need to be able to retrieve smallish chunks of the data back in the typical case quickish.
In turn, that constrains how you group data, and how you allocate ECC bits. As a seller of disks, it's probably how you decide if you should stick a 3TB, 4TB or 5TB sticker on the box too.
The end result will be an algorithm that, based on factory calibration and recent measurements decides how to map data to physical areas of the disk. One head giving a weaker signal lately? Well we're gonna assign fewer data bits there.
And when reading data back there are also decisions to be made. For example, when we read a bunch of data but the ECC recovery fails due to too much noise, should we re-read the same data and combine the reads to get more signal and less noise? Should we shake the head a little and see if we can re-read and get better alignment next time? Should we go and read a few tens of megabytes of other data and see if we can recover the data from a higher hierarchical level ECC/erasure coding scheme?
Discovering Hard Disk Physical Geometry Through Microbenchmarking - https://news.ycombinator.com/item?id=21492345 - Nov 2019 (18 comments)