X
97285

The tide rises for second-wave flash storage vendors

June 13 2019
by Tim Stammers


Introduction


Over the last two to three years, a small group of well-funded startups has brought to market all-flash storage systems that are significantly different from existing systems because they were purpose-designed to maximize NVMe flash performance. This is a second wave of all-flash storage, which abandons the mostly disk-era architectures that powered the first wave of all-flash systems that emerged around 2013. Collectively, this group of second-wave startups has not yet achieved large sales numbers, but we estimate that the total number of production customers exceeds 70, with that number including seven-figure deals and petabyte deployments. The wide range of target applications spans analytics, machine learning, genomics, video processing and transactional processing.

The 451 Take

We have advised in the past that this sector will take time to gain traction, and some of its participants have also shared this view. Now, however, there are strengthening signs of take-up. While the applications are wide ranging, the most common uses for the high-speed second-wave AFAs are for analytics and machine learning. As those two workloads continue to grow, so should take-up of the storage systems sold by the vendors named in this report. The range of architectures they embody is surprisingly wide. Because the IT industry has never ceased demanding more processing power, which in turn demands faster access to data and higher-performance storage, we expect all of the approaches taken here to be adopted over time.

A second wave of AFAs, purpose-built for NVMe


The large majority of current all-flash arrays (AFAs) are based on internal architectures that were devised during the disk era. Indeed, most current AFAs are modified versions of what were originally disk arrays. None of this has hindered take-up, and instead has been helpful by providing continuity for both suppliers and customers. The majority of currently shipping AFAs are also still powered by flash drives based on the SAS protocol – a SCSI-based interface also originating in the disk era.

The industry is in the process of switching from SAS – and the SATA protocol more commonly used in servers – to the NVMe protocol, which was designed specifically for flash and storage-class memories such as Optane. NVMe significantly increases the performance of flash drives and AFAs, and when it is also used in its NVMe-oF networked variant, it boosts the performance of storage networks. Second-wave startups argue that the benefits of NVMe and NVMe-oF cannot be fully exploited by current storage systems that have been converted to use the new protocols. This is because the latencies imposed by the main controllers in their conventional, pre-NVMe architectures are too great, and mask the benefits of NVMe and NVMe-oF, according to the second-wave startups.

451 Research is tracking seven startups that have developed all-flash storage systems that they describe as purpose-designed for NVMe (see below). The architectures of the systems they are selling are varied, and cover a spectrum ranging from systems based on commodity hardware and sold as software-only at one end, to stripped down systems based on specialist hardware with no software or data services at the other end. These systems can also be powered entirely by Intel Optane NVMe drives, as an alternative to flash, although sales of such configurations are rare.

Startups with all-flash storage purpose-built for NVMe Startups with all-flash storage purpose-built for NVMe

Name

GA announced

VC to date

Apeiron Data

Early 2017

$10m

Attala

Mid-2017

Not disclosed

E8 Storage

Late 2016

$18m

Excelero

Early 2017

$35m

EXTEN

Mid-2018

$40m

Pavilion Data Systems

Early 2018

$33m

Vexata

Early 2017

$54m

Applications


Naturally enough, the second-wave startups' sales pitches have centered on performance-sensitive applications, such as analytics, machine learning, genomics, video processing, financial trading and conventional but performance-sensitive transactional databases.

In many cases, those applications are based on server-installed flash or direct-attached storage (DAS), since that form of storage is faster than centralized or shared SAN storage, even when the latter comprises an AFA. However, DAS suffers a number of disadvantages affecting both capex and opex, which is why, in enterprise datacenters, it is used only for a minority of performance-sensitive applications. Because purpose-designed NVMe storage systems are faster than conventional AFAs, they can replace DAS while meeting performance requirements, and in the process reduce costs and increase operational flexibility. All of the systems produced by the vendors listed above provide only block-level access to data, and are frequently used with high-performance file systems such as open source Lustre and BeeGFS, or IBM Spectrum Scale.

Sales progress


This is not a mainstream market, but neither is it a niche – the target customers include hyperscale cloud service providers, web-scale operators and large enterprises, resulting in sizeable deals. More than one member of the group has claimed sales to individual customers running into petabytes of capacity and seven-figure prices. However, public disclosure of sales progress is a very sensitive issue for startups.

The companies in the group were not prepared to reveal exact sales numbers, and some were not prepared to reveal any such numbers at all. Three of the group say they have passed well into double-digit customer counts. 451 Research estimates the total number of customers won collectively by the seven startups to be 70-120 customers. Twenty-six (26) customers have been named, and a handful of case studies have been written about customers that did not want to be named. Companies in multiple industries can be reluctant to become public references. This sector includes many financial customers, which are highly resistant to revealing the names of their technology suppliers.

Venture funding has been economical. There are no repeats of what (in some cases) was very generous funding for the first wave of all-flash startups about six years, when, for example, Violin Memory and Pure Storage raised $254m and $470m, respectively, before making their IPOs. Investor enthusiasm for storage has waned, and the second-wave startups are being far more frugal. Additionally, they are not aiming at a high-volume sector requiring heavy spending on sales and marketing.

Competition


One of the biggest obstacles to sales for the second-wave startups is customer inertia, or the willingness of IT organizations to continue with existing DAS or local server-installed storage, rather than investing in a switch to shared SAN storage that would help reduce opex and drive up flash storage utilization rates.

Another source of competition is the incumbent suppliers that have converted their previously SAS-powered AFAs to NVMe. By the end of 2019, this is likely to be all of the major storage suppliers. All of the incumbents will be targeting the same applications as the second-wave flash storage startups. However, these converted systems are based on conventional architectures, and none of the major vendors has yet shown interest in exploring alternatives. The exception was pre-merger EMC with the DSSD D5 product launched in 2016. Despite the fact that the D5 was taken off the market less than a year later because of what Dell EMC said was limited market demand, we believe its brief existence was a good omen for the second wave of AFAs. The D5 was developed by a startup called DSSD, which was acquired by EMC, highlighting the possibility that, when demand develops, acquisitions of the current startups are very possible.

Indirectly, other competition comes from vendors that are also attempting to improve overall application performance by providing faster access to data – but are doing so with different approaches, such as computational storage or the extension of DRAM using Intel's Optane NVDIMMs.