Backs Up Against The Memory And I/O Walls

The memory and I/O wall are arguably the two most intractable bottlenecks in the modern datacenter.  And while those limitations are a source of consternation for cloud and enterprise users alike, for a company like Micron Technology, whose products are focused almost exclusively on these two kinds of devices, those bottlenecks can be seen as market opportunities.

Of late, the company has been particularly focused on expanding its footprint in storage. Just recently, Micron launched its initial 3D XPoint SSD, the X100, which is the first alternative to Intel’s Optane SSD. The X100 boasts 9 GB/sec of I/O, with latencies an order of magnitude better than NAND flash SSDs. Significantly, the X100 uses the NVM-Express interface, which is well on its way to becoming the de facto standard for solid state storage in the datacenter.

In conjunction with the X100 unveiling, the company simultaneously announced the 7300 SSD series, another NVM-Express product, this one aimed squarely at mainstream cloud and enterprise users. In fact, it’s the first mainstream offering in this bracket, filling a gap between Micron’s top-of-the-line 9300 NVM-Express SSD and the just-announced 5300 SATA SSD. By the way, both the 7300 and the 5300 employ Micron’s latest and greatest 96-layer TLC NAND technology. It represents the third generation of the company’s 3D flash memory, following the 32-layer and 64-layer designs.

The 7300 is destined to be the star, though, using its NVM-Express smarts to deliver up to 3 GB/sec of read throughput (with average latency of 90 microseconds) and 1 GB/sec of write throughput (with average latency of 25 microseconds).  Capacities range from 960 GB all the way up to 6.4 TB.  As a mainstream offering, it covers a lot of application territory, from transaction processing to analytics to machine learning.  And as an NVM-Express device, it will find a suitable home in nearly any server of recent vintage.

Micron’s enthusiasm for NVM-Express is completely understandable, given the large following the technology now enjoys in the datacenter. According to semiconductor market research firm Forward Insights, the enterprise SSD market share of NVM-Express devices is expected to hit 60 percent this year, with SAS and SATA SSDs, together, making up the remaining 40 percent. That would reverse the 40/60 split of NVM-Express to non-NVM-Express market share in 2018. Regardless of the exact numbers, we are seeing a pretty amazing adoption curve when you consider that NVM-Express has only been around since 2015.

“Now I personally don’t believe that datacenter will go to all-flash anytime soon,” says Roger Peene, the vice president of marketing for the Storage Business Unit at Micron. “But the penetration and footprint is absolutely growing at an astounding rate, predominantly at the expense of hard drives.”

This growth is particularly apparent in the hyperscale space, where Peene tells us they are seeing almost 100 percent transition from solid state SATA to NVM-Express. That largely has to do with the better SSD performance made possible by the PCIe-based NVM-Express standard. “Putting very expensive flash behind the SATA interface, which limits its throughput, just doesn’t sense,” he explains. “So, on SATA you’re limited to 550 MB/sec.  On NVM-Express, you can scale up to 6 GB/sec and beyond.”

That level of performance has opened up new ways to use SSD technology. One example of that, he says, is the Zoned Namespace” (ZNS), a proposed standard for NVM-Express devices that allows an application to write directly to NAND, minimizing latency and enabling better use of other SSD resources.

Novel solid state form factors are being trialed as well, one of the more talked about ones being the Enterprise & Datacenter SSD Form Factor (EDSFF). The longer version is in the form of a ruler, while the short version is a quarter that size. It enables designers to pack a lot more SSD capacity into a server than before – up to 32 drives in a 1U space.

Again, this is especially attractive to hyperscalers, where the density and cost-capacity efficiencies are particularly relevant. Peene says that the EDSFF form factor, along with the decreasing cost of NAND and increasing densities will enable drives of 16 TB, 32 TB, or even 64 TB.

Peene says the depth and breadth of the hyperscale environment enables those companies involved in it to vertically integrate their entire infrastructure, which makes it easier for them to experiment with whatever hardware they deem worthy of their business goals.

“The enterprise market doesn’t have the luxury of developing those vertically integrated solutions,” he explains. “What we need to do is find those areas where we can create industry standards and the building block necessary to for them to take advantage of the technology very similarly to the hyperscalers.”

According to Jason Echols, who leads the technical marketing team at Micron, one of the most important factors propelling NVM-Express now is the maturity of the ecosystem. He recalls when the technology was introduced just a handful of years ago, there was little software available to take advantage of the increased performance and capability. He says if you were employing NVM-Express in a caching tier, you would see maybe a 10 percent or 12 percent performance boost in applications like SAS or SQL, which was nice enough, but left a lot of its potential on the table.

Today, with more software written explicitly to support NVM-Express, performance can often be boosted by a factor of two or more. For example, Echols says they’re seeing a quadrupling in performance running SAS on Red Hat when backed by NVM-Express hardware. “Now, you can not only use it for the cache tier, but the capacity tier as well,” he explains. “Both of those allow you to do a lot more work within a single node.”

All of which goes to explain why Micron is investing in NVM-Express products like the 7300 and the X100. That said, the SSD market is not monolithic and they believe that demand for solid state SATA devices will continue for some time. Which is the rationale behind the launch of the new 5300 product.

“We tend to over-estimate the front-end ramps of new technologies, “Peene reminds us, “and we tend to underestimate the back-end tails, especially when we’re talking about traditional enterprise.”