Enmotus Blog

Storage Visions 2017

Posted by Jim O'Reilly on Jan 18, 2017 2:22:42 PM
Find me on:

Here it is. A new year opens up in front of us. This one is going to be lively and storage is no exception. In fact, 2017 should see some real fireworks as we break away from old approaches and move on to some new technologies and software.

These are the highlights going into 2017:

  1. NAND prices have risen some as foundry capacity has been taken off line for conversion to the much denser 3D NAND process. We can expect this increase to hold up till around mid-year, when the conversions will mostly be over, then prices will drop back to their low point in 2016. After that, new foundry capacity coming on stream will keep us on the declining price curve.
  2. With 3D NAND, bulk capacity SSDs will reach as much as 100TB in a 2.5 inch form factor. One incidental consequence of this is that servers and hyper-converged appliances will shrink dramatically in size, with as much as a 50 percent footprint reduction.
  3. NVMe will take over from SAS as the primary storage interface within the server. SATA will struggle on for a while, but the recognition that all of these cost the same to implement will eventually unify in-server storage on NVMe.
  4. NVMe drives will appear in the pint-sized M2 form-factor (This isn’t fair! It happened last quarter!) again opening up server shrinking further.
  5. Micron may beat Intel to market with a fast X-Point product. How much faster than NAND? Still a lot of hype but definitely an improvement – Micron claims 10 microsecond latencies against the 150 typical of NAND.
  6. We’ll see ReRAM and maybe X-Point NVDIMMs with byte-addressable memory on board. This is close to persistent DRAM, a dream of many system designers. There are lots of software issues to lay to rest, from compilers to operating systems recovery routines even before app tuning can really begin, but the next generation of coders will look back in wonder at 4KB block-IO and all the software gymnastics it implies. Byte-addressable means that a single CPU command can read or write data with almost no latency – the I/O stack takes thousands of instructions and, just as important, uses much more precious bus bandwidth.
  7. Networks for storage are going to speed up. 25GbE links will hit mainstream in 2017, with quads running at 100 GbE. Motherboard support will be demanded by the mega-CSPs, followed by the rest of us mere mortals, so introduction will be fast.
  8. Hyper-converged systems will be white-hot in 2017. The approach makes a great deal of sense. Adding RDMA to the 25GbE links will address much of the speed problem in early systems.
  9. NVME over RDMA over Fabrics will be the hottest network sharing technology in 2017. Is it overkill….core counts on CPUs are jumping, while NVDIMM is effectively multiplying DRAM by 5x in capacity. The short answer is NO!
  10. Server core technology is about to take a step up. CPU/memory modules based on some variant of Hybrid Memory Cube are a hot topic at IT’s leading edge right now. These will add 16 or 32 GB of close-in high-bandwidth DRAM to each CPU, effectively creating an L4 cache. But there’s more! The same technology is being explored for the remaining bulk of DRAM, plus any NVDIMM memory. It involves many independent serial channels to the memory, which coincidentally run at much lower power than today’s approach.
  11. Servers will host containers, rather than virtual machines, with a changeover ongoing through the year. With 3 to 5x the instance count of hypervisors and a much faster startup time, containers are a hot software topic. (This incidentally drives IO demand up, making the hyper-converged approach more attractive.)
  12. Object storage will match or surpass the other approaches in performance in 2017. The reason is massive parallelism in accesses, coupled with much improved code (that finally figures out SSDs!) Inter-node links will have the same demands as hyper-converged systems, while the average appliance will look identical to the HCI box. The net is that object storage will become software-defined storage as well, with Ceph running in virtual instances.
  13. Object storage will really embrace deduplication and compression of data (Ceph doesn’t yet have either). This will make a typical appliance by end 2017 appear to hold more than 2PB of effective storage (just 4 of those 100TB SSDs, compressed)
  14. Bleeding edge work on accelerating encryption, compression, deduplication and erasure coding will surface in 2017. I’ve seen it run, and it is blindingly fast. This will close many holes in how we manage data for smallest footprint, best data integrity and greatest security.
  15. The use of extended metadata, couple with services running in virtual instances (collectively. Software-defined storage) will explode in 2017. The metadata approach means processing and managing storage will be data-driven and highly automated. Expect lots of very powerful features and more startups than we can find names for!
  16. Storage analytics will become important as containers, SDS and hybrid clouds make storage configurations and workflows much more dynamic. Automating these areas will be a high priority of any good SDI orchestration suite.
  17. Databases will get even faster as in-memory approaches are supplemented by NVDIMM and faster LAN connections. The war between Oracle and SAP will heat up, and Larry Ellison will thoroughly enjoy himself!
  18. More mergers between traditional vendors will occur. Dell and EMC recognized that hardware is a shrinking revenue source, as size, storage capacity and performance improves, Chinese competition from white-boxes heats up and storage becomes a commodity. Others will seek the same solution of merging. This is already in play … the dance cards have limited options. I’m not going to share my choices, since that would be too speculative and anyway, why spoil the surprises, but I’ll tell you if I won next year.
  19. Realpolitik is raising specters of the 1970’s. A tariff on Chinese IT gear will upset apple-carts (no pun, honest!) and create a short-term problem of higher prices. China will retaliate, but the impact won’t be so much in a reciprocal tariff (China does enjoy a huge surplus in trade with us), but in limiting loan availability and otherwise impeding any on-shoring we try to do. The consequences may include many enterprises (and AWS/Google/Azure) moving IT out of the US to much cheaper countries over the next 4 years, for example.


I’ve probably missed a few, and each of these issues needs more than a paragraph to explain in detail, but I think this hits the big ones for the year, at least from the viewpoint of January.

Have a Happy New Year!


Want to See What Your Storage Is Doing?  Free Trial

Topics: NVMe, SSD, Data Center, data anlytics, NVMe over Fibre