All-flash arrays, whether built with fast SSDs or custom flash cards, are the race cars of the data storage world, characterized by IOPS numbers in the millions and fast connections to host systems. A typical top-flight AFA storage system achieves 4 million IOPS and has two or more 40 Gigabit Ethernet interfaces.
The hybrid storage array seems pedestrian in comparison. Typically supporting four to eight SSDs, with HDDs bulking out storage, these arrays are theoretically cheaper than all-flash array storage, given the somewhat higher cost per gigabyte of flash storage.
"Theoretically" is the important word. Price is a complex number in storage, due to the high markups in traditional arrays. This leaves all-flash array storage vendors a bit of wiggle room to take less of a profit margin and market a more attractively priced system. Even so, flash is a bit more expensive as priced to the user.
If that were the whole story, AFA storage would find some resistance to breaking out of its niche as the choice for the best performance. In the high-performance computing market, on the trading floor or in engineering simulation, performance is king, and those 4 million IOPS make for an easy sell. But for more mundane work, total cost of ownership matters as much as IOPS, and maybe more.
Digging into the cost issue
The hybrid selling point is that only part of the storage needs to be solid-state. The rest can be cheaper, slower disk drives. All-flash array storage vendors deliver only solid-state storage, but they craftily support auto-tiering data to slower bulk storage arrays (using HDDs), the cloud or object storage. The original selling proposition for all-flash array storage was as a drop-in accelerator to existing SANs, a task that was achieved in a stellar manner.
But, as the pitchman says, "Wait, there's more!" The controller in a storage appliance has a major impact on functionality. One common all-flash array feature, data compression, provides a major advantage over RAID arrays built with limited features. With more than enough raw bandwidth to satisfy connected hosts, AFAs use some of the excess bandwidth to compress data in the background. At mega-IOPS levels, this uses a lot of compute power and IOPS. The typical AFA has enough compute power to drive this need, giving it an advantage over hybrid products.
The result of this compression is that the capacity of the flash and any attached HDD or object storage is effectively multiplied by a factor typically around 5x. This tips the cost equation strongly toward AFAs.
The market impact of this is two-fold. The AFA and hybrid array are rapidly eating up the traditional RAID array market. For example, IDC reported that AFA storage accounted for 17% of the external storage market in Q2 2016 in Western Europe, while hybrid arrays came in at 46%.
Die is cast for the future
As we look into the future, the cost of flash will improve. The preferred flash technology by mid-2017 will be 3D NAND, with a resulting multiplication of the per-die capacity. Even though the processing cost of an individual wafer or die increases considerably, the capacity increase outweighs this by a large factor, and, consequently, flash pricing will drop rapidly in the second half of 2017. From late 2017 onward, we will also see a huge increase in flash wafer production as new foundries come online.
The result of this die cost activity will be a price structure that favors all-flash arrays over hybrid arrays. HDDs will move toward end of life in 2018, with much larger capacity SSDs, such as the planned 2020 release of the Samsung 100 TB unit described at the 2016 Flash Memory Summit, replacing low-capacity, enterprise nearline drives stuck in the 10 TB to 15 TB range. From this perspective, AFA storage has won the battle.
However, the characteristics of AFA storage units are changing, driven by the phenomenal performance of the latest generations of top-end flash drives. We now have single SSDs that reach 10 million IOPS, and we can expect these numbers to continue to increase. I/O bandwidth requires a more compact system design, with just a few of these drives (or their equivalents in proprietary packages) mounted in a single appliance.
This type of unit is already seen in the hyper-converged systems market, driven by powerful server engines rather than inexpensive controllers. With the small footprint of SSDs (those 100 TB drives are 2.5-inch units) the hyper-converged infrastructure system has room for a host of screamingly fast primary drives and the bulk storage drives needed to complement them, all with data compression. Such a 2U appliance in 2018 will have 200 TB of primary storage and 2 petabytes of secondary storage, based on a 5:1 compression ratio.
In essence, this is the same design as an all-flash array in that timeframe, though the AFA may still have dual controller architecture coupled with proprietary flash drives in some cases. This leaves little room for a hybrid array.
The question of interfaces needs to be addressed in all of this activity. Today's AFA storage systems support Fibre Channel (FC), due in part to their original use as SAN accelerators. To handle the much higher data rates to come by the end of 2017, network improvement is essential. This involves moving to remote direct memory access (RDMA), which will resolve into NVM Express over Fabrics, a new extension of the nonvolatile memory express access protocol currently used in top-end SSDs.
We are at a fork in the road on what the "Fabric" will be. RDMA is mainstream in the Ethernet and InfiniBand market, and it looks like Ethernet will be the clear winner going forward. That places FC in the gunsights, with a performance and market momentum disadvantage and no RDMA experience.
This suggests that all-flash array storage in 2018 will break free of the FC SAN altogether, which fits with the evolving software-defined storage model. Taken together, all-flash storage appliances will likely take over the market in just a few years' time.
Which is better, hybrid or all-flash?
Learn if all-flash arrays are right for you
All-flash storage could be the norm in Europe by 2020