By Andy Watson, CTO, WekaIO

Storage Class Memory, Redefining "Tier 0"

"The faster you go, the shorter you are."
- Albert Einstein

My prediction for 2020 is that we'll be able to go a whole lot faster than ever before, at least in terms of data access. So, given that Einstein was right about so many things, maybe we should prepare ourselves to get a great deal shorter.

In 2020, persistent memory will begin to deliver on a long-elusive promise. True SCM (Storage Class Memory) has been bandied about for years, with nothing but trickle-down limited-functionality Intel® Optane^TM placeholders available on PCIe like a teaser for a movie with no fixed release date. But ever since at long last Intel delivered Optane DC Persistent Memory in mid-2019, the industry's long-held breath could finally be exhaled.

Now we can experience Optane-based data access with latencies below 1 microsecond.  (Actually, down around 0.35 microseconds.) That's about 1000x faster than the sub-millisecond latencies quoted for ordinary NAND-based flash storage alternatives, and 20x faster than a PCI-based Optane SSD.

Some of you might be getting stuck on my singling out Intel's Optane.  Well, okay, it won't always be only Optane; but for now, Intel is shipping and its rivals are lagging. How long will it take for Toshiba's XL-Flash, Samsung's Z-SSD, or any other contender to evolve into a meaningful frontside-bus alternative? Realistically, the alternatives still appear to be more hypothetical than realistic at this stage. But Intel's Optane is available, it works, and it's compatible with Intel's Cascade Lake CPUs.

Some readers may be aware of other technology contenders in the SCM race. MRAM (Magneto-resistive Random Access Memory) has been around for more than 10 years and is both faster than Optane and more durable (capable of a greater number of write cycles before wearing out). But its storage capacities are tiny by comparison, making it irrelevant to the sea change I see coming in 2020. And although other emerging technologies such as ReRAM (Resistive RAM) and PCRAM (Phase Change RAM) also have potential, none of the alternatives have thusfar yielded anything commercially relevant. (N.B., Optane is based on technology related to PCRAM.)

Regardless of the way it's implemented at the chip level, for years we've been warned that SCM will be useless until somebody with a monstrously big brain figures out how to make "Memory-Centric Computing" happen. Apparently, that's not the case. Companies like MemVerge, Levyx, and others are already making their software available to customers to explore how to use SCM. They're making it possible to run existing applications on SCM without any code changes whether harnessing this new sub-microsecond layer as extended memory or while accessing it as insanely fast persistent storage.

My prediction for 2020 is that next-generation software will move beyond the beta-release stage to enable enterprise customers to unlock tremendous business value with SCM - without waiting for major revisions to the von Neumann computing model. The monstrously big brains are welcome to continue working on memory-centric computing, but there's no reason to wait.

This will take two forms:

1. using Optane to extend the available memory footprint into the hundreds of terabytes (via clustering); and
2. using the persistent memory capability of Optane to offer data storage with unprecedented sub-microsecond latency characteristics.
   

In-Memory Databases like SAP Hana, Apache Ignite, Kinetica, and even good old-fashioned Memcached - or any other memory-intensive applications - will enjoy hitherto-impossibly-large footprints to do hitherto-impossible things at unprecedented scale. Checkpointing long-running tasks will become virtually instantaneous (as will recovering from those checkpoints). So-called "big data" analytics will be accelerated dramatically even while running at larger scale than has ever been previously possible. All of these developments will make it more realistic and cost-effective to apply AI to a wider range of practical problems.

As SCM is adopted and deployed in production, you can bet that the performance expectations for the next layer adjacent to this new "Tier Zero" will be difficult to satisfy. Although dozens or hundreds of terabytes of SCM might be deployed to hold the working set of data, the size of an entire Data Lake nowadays is typically measured in petabytes. At WekaIO, we anticipate that the performance demands associated with refreshing the SCM-based working set will be daunting. Luckily, our flash-native Tier One solution is widely known as "the world's fastest filesystem" and so we are confident that our filesystem will be the obvious choice as the Tier One storage layer associated with SCM-based Tier Zero.

##

About the Author

Andy Watson has focused on the customer experience of high-performance file storage since 1991. After joining NetApp in 1995 (serving as its second CTO there 1998-2007), Andy led cross-functional teams to integrate multiprotocol file server appliances into increasingly complex environments as diverse as Oracle dbms and Rocketmail (the precursor to all webmail), or spanning the spectrum from SAP production environments to seismic analysis applications for the energy sector. Most recently Andy served as CTO at Minio, a cloud-native open source object storage software company, where he confronted cloud-vs-on-premises decisions and the challenges of digital transformation.

As CTO at WekaIO, Andy actively engages with customers and partners, monitoring emerging trends and bringing actionable feedback to Engineering and Product Management.

Andy has a Bachelor's degree in Physics from Bard College.