X
Tech

What we learned about SSDs in 2015

2015 was the beginning of the end for SSDs in the data center. Why? Because researchers have delved deep into their actual behavior and found multiple problems. Here's what you need to know.
Written by Robin Harris, Contributor

Despite their wide use, SSDs are a young technology, one we're still learning about. Here's a roundup of the best research on SSDs in 2015.

Performance

Researchers at Facebook and Carnegie Mellon checked out both reliability and performance of SSDs. They found that high temps can cause SSDs to throttle back on performance. Slow server? Check SSD temp.

From researchers at SanDisk we learned that the log-structured I/O management built into SSDs is seriously sub-optimal for databases and apps that use log-structured I/O as well - which today is most of them.

But the most startling SSD paper came out of Korea, where researchers concluded:

In this paper, we show through empirical evaluation that performance [Service Level Objectives] cannot be satisfied with current commercial SSDs.

As the paper shows, using an SSD poorly can waste most of its possible performance. And until vendors give users the right controls - for example, pausing garbage collection - SSDs will inevitably fail to reach their full potential.

Finally, the unpredictable latency of SSD-based arrays - often called all-flash arrays - is gaining mind share. The problem: if there are too many writes for an SSD to keep up with, reads have to wait for writes to complete - which can be many milliseconds. Reads taking as long as writes? That's not the performance customers think they are buying.

Reliability

The Facebook paper offered the best data on SSD reliability. Key findings:

  • System write activity correlated with SSD failure, probably because flash writes require a lot of power.
  • SSD unrecoverable read errors are relatively common: 4.2 to 34.1 percent of the SSDs reported uncorrectable errors.
  • SSDs are sensitive to temperature - more so than hard drives.

Replacing flash

Today's SSDs run on NAND flash, which is far from the ideal storage medium. Clunky addressing. Very slow writes. Poor endurance. And it's an analog medium, driving vendors to 3D architectures.

But it's cheap, thanks to widespread consumer use, so engineers have made it work in much more demanding applications. But better alternatives are on the way:

3D Xpoint. Intel and Micron's hastily announced 3D Xpoint and Optane drives are promised for 2016, but I'll believe it when I see it. It is supposed to combine the performance, density, power, non-volatility and cost advantages of all available memory technologies on the market today. The technology is up to 1,000 times faster and has up to 1,000 times greater endurance than NAND, and is 10 times denser than conventional memory.

But as details have continued to trickle out, the future of 3D Xpoint looks less certain. Incompatibilities with current tech, pricing concerns, single sourcing and more are clouding the picture.

Nantero. The company hopes to be the ARM of memory technology, licensing to all comers. Their carbon nanotube memory offers promises like those of 3D Xpoint:

  • Fast as DRAM - with much lower power consumption
  • Unlimited endurance
  • Non-volatility - >1000 years at 85C
  • Picosecond switching smaller feature sizes than flash - down to 5nm

I wrote about Adesto last November, so they aren't 2015 news, but they are another NVM technology that could surprise us one day the way flash did 10 years ago.

The Storage Bits take

SSDs have always been a transitional or bridge technology. There's no way that we'd be using SSDs today if we'd had flash technology in 1957 instead of IBM's RAMAC.

But billions of open SATA ports made for a ready market. And now that the early fears of low endurance have passed, we're ready to move on. NVMe, 3D, TLC and all-flash/no SSD arrays will drive the market in 2016.

Since CPUs aren't getting faster, making storage faster is a big help. We can expect more of that in 2016, along with much lower flash prices.

Comments welcome, as always.

Editorial standards