Computational Storage? Don’t Worry – It Hasn’t Gone Away

Gartner’s hype cycle describes the way technologies progress from birth to adoption. The cycle begins with an ‘innovation trigger’ that soon leads to a ‘peak of expectations’ followed by a steep descent into the ‘trough of disillusionment.’ 

Computational storage (CS) currently appears to be deep in that trough. But in Gartner’s model, the trough is followed by a rapid climb to the ‘slope of enlightenment’ and, implicitly, adoption. Adherents of CS are optimistic that the technology they have been developing for the last several years will soon reach that final stage – and that standards have been the missing element holding backing deployment of the technology.

CS became a very hot topic in IT infrastructure in 2019, soon after the Storage Network Industry Association (SNIA) formed a technical working group (TWG) dedicated to creating CS standards. Given that SNIA is the pre-eminent industry body for data storage vendors, this event was significant. Within months, the TWG had become the fastest-growing and most widely supported group of its kind in SNIA’s history, and CS promptly received industry-wide attention.

The concept was also embraced by hyperscalers. In 2019, CS start-up ScaleFlux said it had won a major deployment at Alibaba, and quoted the Alibaba’s cloud database GM as saying that it was “transforming” its storage infrastructure by modifying application code to hand over processing work to CS devices made by ScaleFlux. The start-up also said that three other web businesses were planning to make the same move. Two years later, Amazon Web Services (AWS) launched a service called AQUA, powered by its Nitro processors. Although AWS’ description of the service doesn’t refer to CS, the hyperscaler says on its website that “AQUA pushes the computation needed to handle [data] reduction and aggregation queries closer to the data. This reduces network traffic [and] offloads work from CPUs.”

Related:Why Computational Storage May Be the Next Big Thing

An Established Architectural Practice

AWS’ explanation of the way hat AQUA works happens to be a concise description of CS, which is not a complicated concept, despite its unwieldy name. At a simplistic level, CS modifies the conventional approach of moving data to where the processing is done, by instead moving some processing to the data. This promises benefits for the performance, cost, and physical footprint of IT infrastructure.

Related:Persistent Memory vs. Computational Storage

If all that sounds familiar, it’s because CS is distributed computing applied to storage. Although distributed computing might sound complex and cutting-edge, it has already been widely applied across IT. The humble RAID card is an example of distributed computing. Elsewhere, distributed computing transformed NVIDIA from a relatively obscure maker of graphics processors into a key enabler of artificial intelligence.

A Broad Range of Potential Applications

The potential benefits of CS apply to workloads ranging from big-data analytics and video processing, through data management services such as data compression that are provided by the CPU controllers in storage systems. This breadth of possible applications helped drive the strong vendor support for CS. Founding members of the SNIA TWG included Arm, Lenovo, Micron Technology, NetApp, Samsung Electronics, SK Hynix, and Xilinx. Smaller companies and start-ups were also founders. Other large vendors such as Cisco Systems, BroadCom, Dell-EMC, HPE, IBM, Intel, Microsoft, and VMware joined later.

With all this happening, what could possibly go wrong? Ironically, the breadth of support may have slowed the development of CS standards, according to SNIA. In January this year, the CS TWG had 50 participating companies and 258 contributing members. Scott Shadley, former co-chair of the TWG and SNIA spokesperson said: “As you increase the number of contributors, you gain their different perspectives and experience – and that can create complications.”

Absence of standard has been why CS supporters implicitly acknowledge that there has been no widespread uptake yet by enterprises of the CS products on offer from Samsung and start-ups. Although the first of those products was launched as long ago as 2017, CS vendors are tight-lipped about shipment numbers, and in some cases will not even discuss types of buyers or applications. That alone raises doubts about take-up. In addition, one of the leading lights of the sector, start-up NGD Systems, was recently forced to close its doors because of lack of sales.  And despite its successes at Alibaba and elsewhere, ScaleFlux as another leader has heavily changed the way it promotes its CS devices – because, as it freely acknowledges, the concept has been a tough sell.

There are different types of CS devices. ScaleFlux’ CS devices are flash drives that can take on computational work, but the company’s experience reads across to other CS products. Using ScaleFlux drives to directly accelerate applications requires software to be modified and for most IT organizations “the idea of programming a drive is slightly terrifying,” according to Brendan Wolfe, ScaleFlux senior director of marketing.

“Many customers want a better SSD, but they do not want an exotic new technology called computational storage,” he said.. “The perception of risk and complexity is just too much to overcome.” As a result, Scaleflux has de-emphasized the use of its drives to accelerate applications in the way that Alibaba did. Instead, it is pitching the devices as a means of boosting performance by handling services that would normally be completed by storage system CPU controllers, such as compression and encryption. While that is still CS, it is a significant change.

Wolfe says Scaleflux can easily demonstrate that using its drives to handle data compression can double overall application performance. Compression is handled by the CPU controllers in virtually all storage systems, and vendors of those systems warn that it can heavily reduce performance. However, this is not the case for encryption, because almost all enterprise flash drives are self-encrypting. (On this point, ScaleFlux says that as drives become faster, self-encrypting drives will become harder to build, increasing the value of ScaleFlux’ technology. ScaleFlux and other CS devices could also handle developments in other data management services such as protecting data integrity.)

Standards to the Rescue

Removing direct application acceleration and leaving only compression appears to be taking a lot off the table. However, the expectation is that standards will bring the full suite of CS potential back into play and will ultimately lead to application software supporting CS out-of-the-box. “Some ISVs have enabled their products to use CS, but most will wait until they see signs that the market is adopting it. Broad adoption will occur when it’s trivial to do so – with low risk and low complexity,” says Wolfe.

The first installment of the standards being developed by SNIA’s technical working group was issued last year, as the CS Architectural and Programming Model v1.0.  Shadley says this will make application modification easier. “If developers work from a common framework, then porting that work across vendors is less complex. Having a common set of building blocks is the key,” he said. 

An API for control of the CS devices is also being developed and is expected to reach version 1.0 during 2023. ScaleFlux senior director Keith McKay says the API will be critical.

“It will become the de facto standard for mediating between application logic and CS computational storage devices across vendors,” McKay said. “If application developers see the storage industry aligned to a common infrastructure - the SNIA model - it will absolutely start pressing on the scales in favor of adoption.”