SSDs, Flash and Why Disks Will Not Go Away
Almost three years ago, I made the seemingly bold statement that SSDs would not replace spinning disks, and the flash vendors of that time told me I was wrong. Three years later, we still have spinning disks. And though the market units are steady to slightly down, the number of exabytes shipped is steadily increasing.
Flash vendors continue to repeat their mantra that spinning disk is dead. Yet the amount of storage necessary to meet enterprise storage requirements cannot be made by the current or even planned flash fabrication facilities. Add to this the fact that most flash storage vendors are either losing money or barely profitable.
I, for one, am tired of the hype. I figured I would use all my diplomatic skills to bury these inane ideas once and for all.
In this article, I am only going to talk about external SSDs, as PCIe SSDs do not allow data sharing and combined access. Though they have higher performance I/O and and are great for specialized applications, I do not think they provide a general-purpose market solution. Organizations want to share large amounts of data, and PCIe SSDs are not designed to allow that to happen.
Before I begin my rant, let me say that I have been using SSDs for almost 30 years for high performance, I/O, file system caching, file system metadata and high-performance interactive systems. I am in no way, shape or form anti-SSD. I have a pretty good understanding of what SSDs are good for, what they are not good for and where the market is going.
I see are four reasons why SSDs are not going to take over the world:
- Volume and cost
- Infrastructure and vendor count
Volume and Cost
The most recent statistics I have seen say that 13 percent of the NAND market is SSDs. Adding PCs and tablets gets to you up to 25 percent of the market.
In the first quarter of this year, the disk industry shipped an estimated 109 exabytes of storage, but the NAND industry shipped only about one tenth of that. This means that if the NAND industry wants to kill hard drives, they need to build about $700 billion in new plants to increase the capacity and sell their products competitively at hard drive prices.
Guess what? That is just not going to happen. The risks are far too high for the companies making the investments. The pricing model for NAND does not work on a per GB basis compared to hard drives, and this is not going to change no matter what the industry pundits said back in 2007 and 2008.
Almost every SSD, from the highest-priced enterprise offering to the lowest-priced consumer product has a performance profile that performs slower writes than reads. For some enterprise SSDs, the difference are small for streaming I/O, but they still can support fewer write IOPS than read IOPs.
I looked at the fastest enterprise SSDs I could find on the internet, the Stec-Inc s800 series. Here are the performance specs for the s840 (third column on the chart) and the s842 (fourth column on the chart, broken out by size). Across the board, the write performance is quite a bit lower than the read performance.
The specs for a Seagate enterprise SSD look a bit different, as the streaming read and write performance are the same.
Now let’s look at the streaming performance of hard disk drives. First, the Seagate 2.5 inch 15K enterprise SAS drive, which has an average streaming performance of 176 MB/sec with about the same reliability as the two SSDs listed above. For more density and lower performance per GB, the 4 TB Seagate drive streams at about 138 MB/sec.
Everyone knows that SSDs provide far better performance and even far better price performance for random I/O (IOPS problems). But SSDs do not provide better price performance for streaming I/O.
There are lots of streaming I/O problems, for example video capture and playback. Big data capture for analytics is also often streaming I/O. SSDs are not cost effective for these types of problems. Every problem is not an IOPS problem.