Thursday, November 19, 2020

Storage Media Update

My last post on storage media was After A Decade, HAMR Is Still Nearly Here back in July. Below the fold, I look at some of the developments since then.

First, courtesy of Tom Coughlin, three items based on graphs from the latest issue of his fascinating newsletter.

Kryder's law still good

This graph shows that, despite the pessimism I have repeatedly expressed since 2012's Storage Will Be A Lot Less Free Than It Used To Be, the hard disk industry has managed to keep reducing the cost per byte of their products.

Note that the graph is in three sections, initially dropping rapidly, then flat due to the 2011 floods in Thailand, then dropping much more slowly. As I showed back in 2012, the exact Kryder rate, especially in the early years, has a big effect on the cost of preserving data for the long term (see here). So the slowing after the floods has significantly increased the cost of storing data for the long-term above what would have been expected before the floods.

Flash still can't kill HDDs

This graph shows that the result of the hard disk vendors' efforts is that, despite the technological advances by the flash industry (see below), the vast majority of bytes shipped continues to be in the form of hard disk.

I'm somewhat skeptical of Coughlin's projections for rapid increases in total exabytes shipped over the next few years, which would please the good Dr. Pangloss. Although HAMR and MAMR should allow significant increases in disk areal density, and advances in 3D flash should increase bits per wafer, I doubt that these will drive total exabytes as fast as Coughlin projects.

Approaching the limits

A major reason for my continuing (if perhaps futile) pessimism about Kryder's Law is that the closer you get to the physical limits of your technology, the slower and more expensive it is to make further progress. This graph shows that, for about the last four years in the hard disk industry, staying on the Kryder's Law curve has had no help from making the bits smaller. They have had to do it by other cost reductions.

Technologies progress in S-curves, as shown in Dave Anderson's 2009 slide. The areal density graph clearly shows that the current disk technology is at the top of its S-curve. The replacement technologies, HAMR and MAMR, have still not impacted the volume market after being imminent for a decade.

Dave's graph makes it look like the effect of the series of S-curves is a straight-line graph of progress. But the recent S-curves are much slower than earlier ones, so the overall graph has become S-shaped.

Micron improves 3D NAND

Jim Salter reports in Micron announces new 3D NAND process—denser, faster, less expensive that:
On Monday, memory and storage vendor Micron announced that its new 176-layer 3D NAND (the storage medium underlying most SSDs) process is in production and has begun shipping to customers. The new technology should offer higher storage densities and write endurance, better performance, and lower costs.
Building a chip with 176 layers is a truly remarkable feat, a 37.5% improvement over their previous 128-layer product.

There are two different technologies for making the cells of flash memory, floating-gate and charge-trap. Micron's earlier 96-layer technology used floating-gate, their 128- and 176-layer replacement gate (RG) technologies use charge-trap. Charge-trap:
is a type of floating-gate MOSFET memory technology, but differs from the conventional floating-gate technology in that it uses a silicon nitride film to store electrons rather than the doped polycrystalline silicon typical of a floating-gate structure. This approach allows memory manufacturers to reduce manufacturing costs five ways:
  • Fewer process steps are required to form a charge storage node
  • Smaller process geometries can be used (therefore reducing chip size and cost)
  • Multiple bits can be stored on a single flash memory cell.
  • Improved reliability
  • Higher yield since the charge trap is less susceptible to point defects in the tunnel oxide layer
Micron describes their RG technology in a white paper entitled Micron Transitions to Next-Generation 3D NAND Replacement-Gate Technology. They cite the following advantages, beyond the obvious increase in areal density:
  • Reduced capacitance between the cells in a stack, which allows simpler, faster algorithms in the flash controller's write path implementing fewer, sharper voltage pulses to program the cell. This reduces write latency and increases write bandwidth.
  • The use of metal rather than polysilicon for the NAND control gate, which reduces resistance and allows the write voltage pulse to ramp faster. Again, this reduces write latency and increases write bandwidth.
  • Shorter write pulses mean that:
    The strength and time of electric fields applied to the cell material and other NAND structures relate directly to the endurance of the NAND storage cell. The longer the electric field is applied, the more stress is created on the NAND, which reduces endurance.
  • Shorter write pulses use less power, reducing overall power consumption and heat generation
So this technology reduces cost per bit, write latency and bandwidth, and increases write endurance. Salter writes:
If Micron's claims of greatly increased write endurance pan out, it might become possible to replace incredibly expensive SLC (Single Level Cell) enterprise/data center SSDs with much cheaper 3D NAND devices in demanding applications. Meanwhile—assuming no large increase in per-wafer manufacturing cost—the roughly one-third increase in storage density per chip could mean similarly less expensive consumer devices.
Salter concludes:
We don't expect this to be the death knell for traditional hard drives yet. Even in the best possible case—no increase in manufacturing cost whatsoever—this would put the cost per terabyte of TLC NAND somewhere around $85. The cost per TB of conventional hard drives runs about $27, so there's still plenty of air between the two technologies when it comes to price.

Zoned Name Spaces

Anton Shilov's Western Digital's Ultrastar DC ZN540 Is the World's First ZNS SSD starts:
Western Digital is one of the most vocal proponents of the Zoned Namespaces (ZNS) storage initiative, so it is not surprising that the company this week became the first SSD maker to start sampling of a ZNS SSD. When used properly, the Ultrastar DC ZN540 drive can replace up to four conventional SSDs, provide higher performance and improve quality of service (QoS).
The Zoned Name Spaces initiative defines Zoned Storage:
Zoned Storage is a class of storage devices that enables host and storage devices to cooperate to achieve higher storage capacities, increased throughput, and lower latencies. The zoned storage interface is available through the SCSI Zoned Block Commands (ZBC) and Zoned Device ATA Command Set (ZAC) standards for Shingled Magnetic Recording (SMR) hard disks and with the NVMe Zoned Namespaces (ZNS) standard for NVMe Solid State Disks.
The Initiative's Zoned Storage Overview explains:
The zones of zoned storage devices must be written sequentially. Each zone of the device address space has a write pointer that keeps track of the position of the next write. Data in a zone cannot be directly overwritten. The zone must first be erased using a special command (zone reset).
So what is really going on here is an effort to expose the underlying limitations of SMR hard disks and flash storage to drivers and applications. It isn't surprising that Western Digital is "one of the most vocal proponents" of this:
Shingling, which means moving the tracks so close together that writing a track partially overwrites the adjacent track. Very sophisticated signal processing allows the partially overwritten data to be read. Shingled drives come in two forms. WD's drives expose the shingling to the host, requiring the host software to be changed to treat them like append-only media. Seagate's drives are device-managed, with on-board software obscuring the effect of shingling, at the cost of greater variance in performance.
We discovered the "greater variance in performance" when using Seagate's SMR "Archive" drives in A Cost-Effective DIY LOCKSS Box. By exposing the medium to the device driver as a set of append-only, eraseable zones the file system can more closely conform its write operations to the underlying device's capabilities, thus reducing the need to block write traffic while data is moved or erased on the medium. The lack of a standard way to do this has limited adoption of SMR, and has led to greater complexity in SSD firmware.

No comments: