Find out how solid-state drives are going mainstream in the data center.
Given their vast difference in performance, it’s somewhat surprising that flash still hasn’t completely displaced the hard drive in the data center. IT's innate conservativeness is partly to blame, but price barriers and questions of durability have added to the inevitable FUD that hits a new technology set to displace old ways of doing business.
However, as with most major technology waves, there comes a time when the industry moves from timid exploration to wholehearted acceptance. Flash-based products have reached this point, and the migration to an all-solid-state world has begun in earnest.
Seagate has reportedly ceased development of its top-end “enterprise” 15K RPM hard drives. Aimed at the performance market, these drives are very expensive, but for all of that they can’t even come close to solid-state drive speeds and feeds. Equivalent SSDs are NVMe-class units with 400K+ random IOPS. Seagate recently announced drives that achieve 10 GB per second transfer rates. When you look at the 15K drive, its best efforts are 300 IOPS and 150 MBPS, so the writing is on the wall.
You might wonder about SSD durability and price, though. Feverish efforts by flash developers have improved wear life dramatically, so that vendors now promise eight-year life spans. Wear is no longer a problem. Pricing of desktop SSDs has closed in on SATA bulk drives, though there is still a gap. That gap has to be measured against reduced appliance and server counts, which can be very substantial.
More than that, SSD technology is switching over to 3D NAND, which will increase capacity per die by roughly 16x. While wafer completions have dropped during the switchover, causing a temporary increase in prices, next year will see a large increase in foundry capacity to make 3D NAND, which will both lower prices and increase per drive capacity dramatically.
When it comes to capacity per drive, the SSD has just lapped the hard drive. With hard disk development struggling to get past 10 TB per drive, a number of SSD vendors have announced 100 TB units coming in 2017. There’s no way the HDD can reach that number.
New system architectures round off the story. Flash-based NVDIMMs that increase performance by roughly 4x over even the fastest SSD are just entering mainstream use. New systems architectures, such as Gen-Z, are in the wind, offering ways to expand overall system performance dramatically based on ingenious use of flash.
Let’s take a closer look at all these trends that are propelling flash storage into the mainstream.
(Image: Timofeev Vladimir/Shutterstock)
Performance separates today’s storage from the pre-flash world of slow rotating disks. Flash products can achieve more than 1000x the performance, which, for databases and virtualized systems, is a performance bonanza. Flash arrays, for example, can deliver 4 million IOPS, compared to the couple of thousand IOPS from a RAID array with hard drives.
Streaming performance also will improve with the upcoming 10 gigabyte per second SSDs. Against this the hard drive can deliver perhaps 300 megabytes per second. This performance gap has effectively ended the future of the fastest hard drives, the 15K RPM units, and we can expect the next tier of 10K RPM drives to follow shortly.
It’s interesting that hard drive vendors are now promoting the SSD metric of “drive writes per year” on their hard drives. The numbers look comparable to SSD numbers, which means that HDD durability is no longer an advantage.
Careful controller design and improved flash processes have extended SSD wear life to the point that even inexpensive drives can deliver many years with server workloads. One accidental benefit of 3D NAND is that it is built using larger cell geometry than planar NAND, which improve yield but also increases wear life a great deal. With this improvement, MLC (2 bits per cell) NAND has completely supplanted the original single-level cell (1 bit) architecture and we are homing in on 3-bit triple-level cell (TCL) to replace multi-level cell (MLC) in 2017, with quad-level cell (QLC) for bulk storage.
On the price side, there are two factors in play. First, SSD prices have fallen well below 15K hard drives, at least in distribution and from the more enlightened OEMs. Second, there is a realization that top-line SSDs are often overkill and with the new approach of compact appliances instead of RAID arrays leading to appliance level redundancy, even desktop level SSDs meet the needs of many use cases. In terms of cost, desktop-class SSDs are now closing on in SATA bulk drive territory.
3D NAND has a major impact on die prices, with vertical stacks of 64 and even 128 cells increasing capacity per die. The price/bit benefit analysis is complex, due to the increased number of process steps, but it is still large. The improved durability of 3D NAND also means that we can add bit-per-cell gains to the price/capacity equation, with three- and four-bit per cell supplanting one- and two-bit architectures over the next year.
Planar NAND processing ran into both a yield problem and a wear life issue in 2015. The solution is to back off to a bigger feature size, but stack cell layers on top of each other to increase capacity per die. This brings capacity boosts per die of between 16 and 32 times on a 64-layer 3D NAND device, and additional work to stack the 64-cell blocks means that we’ll see 2x to 4x further capacity boosts in the near term.
2017 promises huge drive capacities and this is a direct result of the move to 3D NAND.
(Image source: Samsung)
Foundry capacity controls the price of SSDs and, while it’s true today that demand is ahead of supply and prices have risen somewhat, we can expect a huge amount of new foundry capacity to come online in 2017, coupled with the return of foundries sidelined for the changeover to 3D NAND. This will increase die supply, while 3D NAND boosts die capacity, resulting in a more than adequate supply level as hard drives finally head to the Computer Museum.
(Image: Arvutistuudio/Wikimedia Commons)
Who would have guessed that we’d be talking up 100 TB 2.5 inch SSDs in 2016? The roadmaps are astounding; we are talking about one to two petabytes in a 1U appliance. Hard drives are literally maxed out. We’ll see maybe one or two more generations, but 16 TB and 20 TB hardly cut it, while fringe ideas of using 5.25 inch form factors miss the point of the tiny SSDs.
But there’s more! Software-defined storage should bring real-time compression and deduplication to the game, effectively increasing capacity 5x to 10x. This is possible with SSDs, due to the high IOPS and streaming rates, but infeasible with just hard drives.
Performance requirements are leading the industry to abandon the traditional operating system file I/O stack for very lean NVMe stacks. In turn, vendors are rethinking system design and moving to connect everything from DRAM and close-coupled flash NVDIMMs to SSDS and clusters of computers using the NVMe approach over a yet-to-be-selected common fabric. This approach boosts I/O, networking and compute power dramatically, so the box count for a given workload will shrink.
These advances are only possible with the performance benefits of flash. It’s clear that by 2020, the data center will be a much-changed environment.