1. There will be a resurgence in shared storage (but FC fades)
SANs provide shared block storage to multiple servers. Lately, there has been a trend away from SANs to direct attached storage, and much speculation about the death of SANs, because:
- a. SAN latencies are too high for use with faster SSD storage.
- b. SAN arrays are expensive and proprietary. They often use purpose-built hardware and premium-priced drives.
- c. SANs are complex to manage. SANs based on the FC standard require unique host bus adapters (HBAs), switches, and cables. They also use non-intuitive storage-centric management.
- d. Alternatives to SANs, called vSANs or virtual SANs, have emerged without the above three disadvantages. vSAN software aggregates and shares direct attached storage across a cluster of compute servers.
However, everything is changing rapidly, and the next 4 paragraphs address each of the above four points.
SANs have traditionally used the SCSI protocol over FC. In my view, future SANs will use the emerging, very low-overhead NVMeoF (NVMe extended to work on fabrics) protocol over Ethernet. With NVMeoF, there is no latency difference between local and remote SAN storage. E8 recently announced such a SAN array supporting 10 Million iops with latencies of 40-100 usecs . This addresses issue (a.) above.
Next, future SAN arrays are being built as software on standard x86 servers (e.g. Hedvig, Formation data systems, Tintri) and there is no premium on drive pricing. This addresses issue (b.) above.
Third, SAN arrays are now providing intuitive app-centric management (e.g. Tintri). This, combined with use of NVMeoF over Ethernet, addresses issue (c.) above.
Finally, virtual SANs have disadvantages. They have limited scalability, the approaches they use for high availability (replication, erasure coding) are inefficient compared to RAID, and they are very dependent on good data locality for performance. This speaks to issue (d.) above.
For the above reasons, I believe shared storage SANs (NVMeoF over Ethernet) will see a resurgence. FC, which has long been synonymous with SANs, will, however, fade away.
2. A new storage hierarchy will emerge – Persistent Memory, Flash, Cloud
The current storage hierarchy is complex. From fastest to slowest it consists of server-attached NVMe Flash, followed by shared storage using write-intensive Flash SSDs, read-intensive SSDs, high-speed disk drives, near-line disk drives, and then finally tape or cloud.
The future storage hierarchy will be much simpler. It will consist of persistent memory, Flash and then Cloud.
Persistent memory is an emerging new layer between DRAM and Flash that will be accessed using memory protocols. It will be non-volatile, have latency close to DRAM (100s of nsecs) and provide 4-5X higher capacity and lower price than DRAMs. Many technologies are vying to be this persistent memory layer including 3D Xpoint from Intel/Micron, Resistive RAM from HPE/Sandisk, TLC PCM from IBM and various forms of NVDIMMs that are DRAM and Flash combinations. It is too early to predict a winner here especially given the performance, durability and date slippages with 3D Xpoint which had looked like the front runner to me. Furthermore, much work remains to be done on the software side with OS and file system support before this new persistent memory layer becomes a reality.
The next layer after Persistent memory will be Flash. With NVMe and NVMeoF, this can be Flash in the server or Flash in a SAN array, there is no performance difference. There will be little need for using disk drives for primary storage and we predict all-flash primary storage. Flash prices are dropping very fast – 15k rpm disk drives will not be available past 2017 and 10k rpm disk drives will not be available beyond 2018. Combined with the lower power, cooling and support needs of Flash, all-flash TCO is superior to hybrid Flash/disk TCO or all-disk TCO.
Finally, with the rapid drop in pricing for cloud storage, we see most future archival and backup storage being in the cloud. Cloud storage is most likely to be based on low cost disk drive technology, however a case can be made that next generation QLC Flash (using 4 bits per cell) based archival storage may have superior TCO. As with persistent memory, I am not predicting a technology winner for cloud storage just yet.
To summarize, we see Flash as the new disk, Cloud as the new tape, and we see a new persistent memory layer emerging between DRAM and Flash.
3. Stand-alone primary storage diminishes in importance
Customers used to (and continue to) select and purchase compute, storage and networking products separately. The dominant way to purchase primary storage continues to be as stand-alone storage. The problem with this 3-tier approach is that it can take several months of integration & testing by the customer before they are able to run their first application on this separately purchased infrastructure.
Integrated systems where compute, storage and networking were purchased together emerged in 2007. First generation integrated systems were called converged infrastructure systems (CI). Since the separate products are pre-tested and pre-integrated by the vendor, this reduces the time it takes a customer to run their first application following a CI purchase. While CI systems have good performance and scalability, they are still too complex to deploy and manage.
Starting in about 2012, customers started to purchase a 2nd generation of integrated system called Hyperconverged or HCI systems. HCI uses clusters of compute servers with DAS storage and virtual SAN software. Unlike CI, which is really 3 pre-tested products sold together, HCI is truly a single product with unified management. They are much easier to deploy and manage than CI systems.
More and more, storage will be sold as part of an integrated system such as converged or hyperconverged or whatever they evolve to become. Due to the compelling ease of use and fast time to value of integrated systems, I expect to see stand-alone storage products diminish in importance and gradually disappear.
4. A new form of Cloud Integrated System will emerge
Many people believe that hyperconverged systems will be the dominant, and perhaps ultimate, form of integrated system for the foreseeable future. According to 451 Research’s latest quarterly Voice of the Enterprise Survey of IT buyers, hyperconverged infrastructure is currently in use at 40% of organizations, and 451 Research analysts expect that number to rise substantially over the next two years. Gartner also predicts that HCI will be the dominant form of integrated system by 2020.
I disagree. I believe there will be other forms of integrated systems beyond HCI. This is because HCI systems, in spite of their many advantages, have too many limitations. These include limited scalability (they rarely scale much beyond 16-32 nodes), inefficient resource consumption (storage and compute are coupled), inefficient HA (replication or erasure-coding versus RAID), and dependence on data locality for good performance.
We cannot predict all the new forms of integrated systems that will emerge. However, we are starting to see a 3rd generation of integrated system that combines the performance and scalability of CI with the ease of use of HCI, while providing full cloud capability (self-service, multi-tenancy, metered service, elasticity, network delivered, hyper-scale). In keeping with my previous predictions, they integrate an all-flash shared storage array as part of the integrated system. An example of such a system is the Cloudistics composable cloud platform. Such integrated systems are different from HCI because they allow independent scaling of compute and storage, and they are different from CI because they are not the integration of three separate products, but are truly created as a single product with unified management. Finally, they are different from both CI and HCI in their ability to provide the full cloud experience.