Will Storage Class Memory Bring The Next IT Epoch?
December 10, 2019
Viewed through the lens of geological, anthropological or cosmological history, an epoch delineates a distinct time period when a significant event took place. For example, the Paleozoic Era denotes a geological period when many of the Earth's plants and animals evolved. Unfortunately, that era ended with the "Great Dying," when dinosaurs along with over 50% of all Earth's life forms perished. Over on the cosmology scale, the Planck epoch represents a period of time starting with the "Big Bang" and ending just 10-43 seconds later. (Nobody said epochs had to be long, just significant.)
Epochs exist in the timeline of IT data centers as well. And we might just be on the verge of another: the Storage Class Memory (SCM) Era. Since the invention of computers, memory and storage have been seen as two different logical computer science concepts. Memory is for active data placement during the computation. It requires byte-addressability and is typically accessed by the CPU’s load/store method with cache-line granularity. It does not require persistence. On the other hand, storage is used to provide data durability. It is typically being read and written in data blocks sized 4-kilobytes or larger. It is designed to survive power failures and, in many cases, component failures. Today, DRAM is the primary media technology for memory functions inside computers, while NAND-flash SSDs and Magnetic HDDs are the primary media technologies for storage. Both the memory market and the storage market are each on the order of $100 billion in size, however all of that is about to change because of SCM.
So what is SCM? It is a new class of media that combines the properties of memory and storage. It has a low enough latency that it can serve as byte-addressable memory and it is non-volatile so that it can survive power cycles. That means it has the potential to serve as both memory and storage, and it removes the boundaries between these two computer science concepts. Boasting 10X bigger capacity and 100X faster access, SCM will not only redefine the data center infrastructure but also programming paradigms.
In Q2 2019, Intel shipped Optane DC Persistent Memory, the world’s first SCM product that can be used as general-purpose memory as well as persistent storage. It is based on 3D XPoint technology, one of several competing technologies in the SCM space. It is expected that by 2022 there will be additional major vendors shipping multiple SCM technologies to customers. By 2025 conservative estimates predict that SCM technologies will displace at least 10% of the DRAM market. In addition, SCM is expected to replace some portion of the high-end NAND-flash market, making it at least a $10 billion new hardware device market, ushering a Great Dying of sorts while spurring a rich evolution in applications.
There is a fundamental shift happening in hardware, and it is disrupting the software space, too. The last time this happened was a decade ago when SSDs entered the enterprise storage space. Based on NAND flash, enterprise-grade SSDs started to appear around 2009. The difference between SSDs and HDDs is not as stark as the difference between SCM and SSDs, however it still sparked a major event in the data storage industry. While many system and software vendors were working to adopt SSD technologies into their existing architectures, there emerged a new class of system called all-flash arrays. Vendors designed their systems from the ground up for this new media and quickly found use cases where they shine. This created a new billion-dollar market and created market value of more than $10 billion.
The change SCM is introducing is bigger than what SSDs did ten years ago because it’s not just a faster storage device but also changes the game for the memory market. To look for an event that is comparable, perhaps we have to go back 20 years when Intel started to introduce the multi-core architecture for its CPUs.
Since the CPU clock couldn’t go much faster, Intel started adding more cores (and hyperthreading) into the CPUs. While it didn’t change the x86 instruction set, existing applications struggled to take full advantage of all these CPU cores, given their threading models weren’t designed for it. As a consequence, the CPUs were not fully utilized in data centers. And so VMware’s virtualization technology was born at the right time. By allowing multiple virtual machines to run on the same server, VMware provided a software platform that allowed the power of the new hardware to be fully utilized, without requiring applications to be rewritten. This swept across most data centers and is now a $10+ billion market. VMware, the first mover and leader in this space, is valued at over $60 billion now.
SCM adoption faces a similar challenge today. To fully take advantage of the persistence and byte-addressability of SCM, existing applications will need to be redesigned to a new persistent memory API. Not all applications are ready to do so. Yet, at the same time, the growth in data-centric applications—including machine learning, artificial intelligence, big data and the Internet of Things—demand bigger memory and faster storage. There is a fundamental need for a new layer of software that can express the power of the underlying SCM without requiring the application rewrites, just like VMware did 20 years ago. And the opportunity is as large!
That is where Memory-Converged Infrastructure (MCI) comes in. MCI combines an SCM hypervisor with the world’s first distributed file system and distributed memory system designed for SCM. In this next IT epoch brought in by SCM, some dinosaurs may struggle to survive. But it will also open up the possibilities for new species. MCI software strives to be such a new animal for this new era!