An interesting guest post on Claus’s Blog (Claus Mikkelsen of HDS) by Ian Vogelesang of HGST provided some technical/economic insights on why specific disk drives are more economically feasible than others.
It’s a bit hard going and more technical than a typical blog post, but it certainly makes a number of interesting points.
- There is an interaction between recording density, performance and $/GB when introducing a new, smaller form factor. Most often drive vendors are trying to maximize GB per drive while IO performance is not as much of a concern. So they most often try to first come out with their densest drive they can in any new form factor. I think this is what we have seen with the SFF disks today, i.e., most vendors came out with 10Krpm drives, leaving their faster drives to the LFF. As recording density for a new technology continues to improve, GB/drive is no longer the driving factor which is when performance rises to the top. At that point then we see the introduction of higher speed drives in a form factor.
- Enterprise SATA drives perform worse than equivalent capacity SAS drives. In HDS’s case there are two reasons for this: 1) For enterprise storage they append ECC plus other LBA integrity checks to each 512 byte block however, SATA doesn’t support anything but 2**n block size, thus multiple IOs are required to read/validate a block and 2) SAS hardware supports a larger tag command queue than SATA and thus, a better optimized IO queue for multiple IO requests.
- Global access density requirements are 600IOPs/TB of storage. This is stated as a matter of fact in the post without any background information but is another key factor driving disk changes.
I would love to know more about that last point 600 IOPS/TB. But there wasn’t much else there. (It seem to me this should have changed over time. It’s certainly worthy of a research study if anybody’s listening out there.)
One other thing I found interesting is a few statements at the end regarding emerging disk recording technology. It seems thermally assisted recording (TAR) is not coming along as fast as everyone in the industry thought it would. As such, the disk industry is considering moving to shingled writes (see my post Sequential Only Disk) which may cause them to abandon random writes.
But there is another solution to non-random writes besides sequential only disk and that is implementing a log structured file for blocks on the disk. Similar to NetApp’s Data ONTAP, where the system supports random writes but actually writes data on disk drives sequentially.
This requires more smarts in the drive controller but it’s nothing like what’s in SSDs today for wear leveling and is a viable alternative. The nice thing about a log structured file on disk, is that there is no need to change any IO drivers as the disk drive continues to support random writes (from the server/storage system perspective) but the drives write sequential on the platter.
I would suspect most drive vendors considering shingled writes are busily working on doing something similar to this and it wouldn’t surprise me to see the next generation disks support shingled writes using an onboard log structured file.
What this will do for read sequential IO is another question entirely.
Luckily, data that is read sequentially is often written sequentially and even with a log structured file layout on disk, will more than likely be positioned close together on a disk platter.