UALink – In May 2024 AMD, Broadcom, Cisco, Google, HPE, Intel, Meta, and Microsoft announced the formation of a group that will form a new industry standard, UALink (Ultra Accelerator Link), to create an ecosystem and provide an alternative to NVIDIA’s NVLink.
UALink creates an open ecosystem for Scale-up connections of many Al accelerators
Effectively communicate between accelerators using an open protocol
Easily expand the number of accelerators in a pod
Provide the performance needed for compute-intensive workloads now and in the future
A scale-up memory semantic fabric has significant advantages. Scale-out is covered by Ultra Ethernet and the industry is aligned behind the Ultra Ethernet Consortium (UEC) to connect multiple pods with the ability to scale multiple pods.
This was picked up by Jim “The SSD Guy” Handy, who heard technology and product veep Scott DeBoer say that Micron makes vertically integrated SSDs which include NAND, some DRAM, and the controller. DeBoer said Micron is using an in-house controller because a portion of it was incorporated into the NAND flash chips themselves.
The controller was split between external and internal functions. Why? It means some NAND controller functions can be performed by the NAND chips inside an SSD – low-level functions internal to a chip. That means they could occur in parallel, offload the SSD controller so it can do other work, and enable the SSD to work faster overall.
Handy said an SSD typically has a controller inside it – Micron’s “external controller” phrase – which links to NAND chips across a relatively narrow bus. Each NAND chip is subdivided into blocks and these are linked by a wide internal bus, which is connected to the narrow bus going to/from the controller.
In a process such as the recovery of partially erased blocks – garbage collection – valid data is read from blocks needing to be recovered and copied to an empty block. When all the data from partially erased blocks has been read into the new block, the old blocks’ contents are deleted and they become empty blocks available for reuse. Handy has a diagram showing this:
Jim Handy diagram.
Such garbage collection is directed by the controller, which reads then writes the data involved in the process. While it is carried out, this internal-to-the-SSD process it is not available to perform its main work, reading and writing data to/from the SSD for its host server or single-user system.
Handy suggests that this could be done by putting appropriate internal controller functions – simple ones – inside the NAND chips. For example, the external controller would decide that garbage collection was needed and, using its metadata map of NAND empty and used chip blocks and data contents, calculate a series of block data reads and writes and subsequent block deletions as before. Then it would tell the NAND chips to do the work themselves, and eliminate the internal SSD IO of data reads and writes to/from the controller. A second Handy diagram illustrates the concept:
Jim Handy diagram.
The controller, for example, tells a NAND chip internal controller function to read valid data from block 0 to block 2, then valid data from block 1 to block 2, then delete block 0 and block 1. The data is transferred across the NAND chip’s wide bus, hence it is faster, and the external controller is not directly involved in the IO at all.
What we have here looks like a potential variant of the processing-in-memory (PIM) idea (think Samsung’s Aquabolt AI processor). We could call it PIN – processing-in-NAND. It is literally compute in storage, albeit very simple compute.
This, Handy says, is only a possibility. He writes: “Micron may be doing something altogether different. The company’s engineers may have chosen more appropriate functions to pull into the NAND. In the end, though, the new NAND functions, whatever they are, probably accelerate the SSD while reducing the controller’s complexity and cost. I would also assume that Micron plans to keep these functions confidential, so that only Micron SSDs can take advantage of them. This would give the company a distinct competitive edge.”
It seems a powerful idea and Handy suggests we should “expect to see this approach adopted by other NAND flash makers in the future.”
We must wait and see what Micron announces as it brings out 233-layer 3D NAND SSDs. The signs we’ll be looking for include better-than-expected SSD performance.
Interview: Diamanti started out as a hyperconverged infrastructure appliance vendor but then switched to supplying Kubernetes lifecycle management software that runs on its Spektra all-flash and bare-metal HCI system as well as other systems.
We spoke to CPO/CTO and EVP Engineering Jaganathan Jeyapaul about some of the issues, including customers being more “thoughtful in their Kubernetes choices”, and more.
Jagnathan Jeyapaul (JJ)
The proposition of Diamanti’s Spektra environment is that businesses need a dedicated environment, on-premises or in the public cloud, within which to run Kubernetes and that it is not just another system app to run inside a virtual machine or general server.
Jeyapaul told us Diamanti’s storage software is purpose built to provide high performance and security for cloud native applications running on Kubernetes, adding that through the use of storage accelerator cards, it achieves “about 1 million IOPS per Kubernetes node.”
Blocks & Files: Kubernetes is becoming table stakes for any storage supplier, either on-premises or in the public cloud. What are your thoughts around Kubernetes and its management being a feature and not a product?
JJ: Kubernetes has become the cloud’s default operating system. Provisioning & managing Kubernetes clusters is not a standalone product anymore, rather it is an expected feature within any higher-level cloud product that manages a group of cloud applications and services running flavors of Kubernetes across multi-clouds and hybrid clouds (the service mesh). Diamanti’s control plane and orchestration product has the core ability to monitor, manage and administer Kubernetes across multi-clouds and hybrid-clouds through telemetry data intelligence.
Blocks & Files diagram of Diamanti’s Spektra
Blocks & Files: Will Kubernetes become a cloud-native orchestration tool for IT workloads in general and then also IT infrastructure in general? What are the pros and cons of this?
JJ: Heavy-duty, data-intensive IT workloads (homegrown stateful apps, 3rd party software ex: Analytics) are typically containerized already, and hence run well within Kubernetes. These heavy-duty IT workloads must rely upon a scale-out architecture to achieve performance at scale. Kubernetes serves as an excellent orchestration tool for the scaled-out IT workload nodes. IT infrastructure similarly benefits through Kubernetes adoption for configuration management, deployment, and lifecycle management of infrastructure components.
However, if enterprises aren’t thoughtful in their Kubernetes choices, they could lock themselves inadvertently into different flavors of Kubernetes, which would reduce the portability of their workloads. Diamanti is a great equalizer in that our Kubernetes-based orchestration platform levels the playing field by providing accelerated performance and security while being 100 percent portable between on-premise and cloud clusters.
Blocks & Files: Does Kubernetes have any relevance to composable infrastructure, and if so, what do you think it is?
JJ: Kubernetes enhances a well-designed composable infrastructure and by sharing similar design principles, Kubernetes-based composable infrastructure provides a complete elastic, low-cost (no over-provisioning) and durable infrastructure for data-intensive, scaled-out stateful applications at scale.
Blocks & Files: How is access to Kubernetes secured?
JJ: Kubernetes distributions are generally very secure (specialist vendors like Diamanti thoroughly scan their distros and are certified). However, there are vulnerabilities in the ways a Kubernetes cluster is deployed & administered that require careful investigation. Kubernetes clusters must follow security best practices & standards for RBAC, secret protection, infrastructure as code, container security and end-point protection.
Blocks & Files: What would you say about the idea that Kubernetes is too low-level for mass enterprise use and an abstraction layer with automated functions needs to be erected over it?
JJ: Kubernetes has been adopted by about 90 percent of enterprises already. It is seen as the default cloud OS and platform, and its plugin-based architecture allows for building customized infrastructures to meet all enterprise workload needs. Kubernetes itself must be treated somewhat as a low-level function and a well-designed control plane abstraction for the management of Kubernetes nodes, plugins, and IT workloads is needed in most cases, e.g., Diamanti Spektra.
Blocks & Files: What will the Kubernetes world and ecosystem look like in 5 years’ time?
JJ: Kubernetes will become ubiquitous and will serve as the portable “runtime” for micro apps/services for the serverless, edge and ambient computing use cases. It will be to micro cloud workloads what Java is to on-premises legacy applications (write once, run anywhere, any cloud, any device).
Interview: Quantum, known best for its flash, disk and tape storage, also sells into the data-intensive world of HPC storage. We spoke to Quantum’s Eric Bassier, senior director, Product and Technical Marketing about tape, cold storage, HPC capacities, and more during its exhibition at ISC High Performance 2022 in Hamburg, Germany.
Blocks & Files: What are the use cases Quantum sees in the high performance computing market?
Eric Bassier
Eric Bassier: We don’t do the primary storage for high performance computing. But we do have a lot of customers, different research laboratories, different organisations in life sciences, research, different firms, where they do use Quantum for the secondary storage. And a lot of those use cases are our StorNext file system, with some kind of a disk cache in front of tape.
Blocks & Files: What does an average quantum secondary storage HPC installation look like, in terms of the disk capacity range they might have and the tape capacity rates they might have?
Eric Bassier: It does depend on the use case. But it in general, it might be 10 to 20 percent on disk, and 80 or even 90 percent on tape.
Blocks & Files: Could you describe a couple of customers?
Eric Bassier: One is the Texas Advanced Computing Center (TACC) talk. They’ve built a centralized archive for their research facility based on StorNext and tape. And then the other public case study – it would represent why we want to be at a show like ISC – is what we’ve done at Genomics England. We’ve actually partnered with Weka.
Blocks & Files: Why is that?
Eric Bassier: As a file system Weka is much more suited for a typical HPC type of workload. StorNext really excels for streaming data, which is why it’s so good for large video files, movie files. Genomics England have 3.6 petabytes of flat storage for for their Weka file system on flash.
That’s where they’re ingesting the data from the genomic sequencers. And they now have over 100 petabytes of our ActiveScale object store. It’s totally their secondary storage. In that case more than 90 percent of their data would be considered secondary storage.
ActiveScale diagram
Blocks & Files: Does that include tape?
Eric Bassier: Although Genomics England is not using tape today, as part of their ActiveScale system we are talking to them about it.
Blocks & Files: I’d imagine that the rate at which they’re accumulating data, they’ll possibly start thinking in terms of the disk ActiveScale archive having colder data on it in parts, and maybe there’s so much of it, that they could offload it to tape.
Eric Bassier: It really is an ideal use case in many ways. A reseller partner of ours in the federal government space, they do a lot with kind of AI and machine learning, the head there has said a lot of data is cold, or inactive. But it’s only inactive temporarily.
And their customers … can’t predict all the time when they’re going to want to bring it back from cold data. In many of those use cases, they’re perfectly happy if it takes five minutes, 20 minutes to get data back from tape. The speed of tape is not a factor for that. And they like the low cost and reliability and also the low power – the green aspect.
Blocks & Files: With you having a focus on secondary storage for HPC market, then I guess you’re thinking that we need to accept data from the primary storage systems quickly and straightforwardly and easily, and we need to ship cold data that’s now warmed up to those primary storage systems in the same way. Is there a workflow aspect to this?
Eric Bassier: Yes. Any type of research is going to have a workflow associated with it. They’re going to have a stage where you have scientists actively analysing [or] working on the data, and then they would move it to less expensive storage, to an archive. Now, I think one thing that Quantum has done, where we have a very, very unique offering, is the way that we’ve integrated tape with ActiveScale. I think that it’s the first time where it’s not a tape gateway.
In other words: we built an object store where you can have a single namespace across disk and tape. And the way that an HPC application would interact with it is use S3, standard to read and write objects to disk. And then use either an S3 Glacier API set to put objects on tape and restore objects from tape, or used what’s called the AWS lifecycle policies, which are part of the S3 standard API set.
There are other other solutions out there. There are gateways to put data on tape. But now you’re talking about different namespaces, different user interfaces, and multiple key management points. What we’ve done with ActiveScale, we think is unique, because it’s the only object store where you can create an object storage system on both disk and tape; you can take advantage of the economics of tape. I think we’ve abstracted the way that an application has to interact with tape in a way that’s better than what anyone’s done in the past.
Blocks & Files: Are there any other advantages to ActiveScale for HPC users?
Eric Bassier: The second, really the key innovation for us, is the way that we do the erasure encoding of the objects on tape – that’s where we have patents. And why that matters is you get much better data durability on tape, and you get much better storage efficiency. Instead of making three copies of a single file where you’ve tripled your tape capacity, we erasure code the object and then we create the parity bits and we striped it over tape and it’s more efficient. The other thing that it unlocks through the way we do erasure encoding is that, most of the time, we can recover an object from cold storage with just a single tape mount. And it turns out that’s been a really difficult technical challenge to enable this concept for many, many years.
Blocks & Files: I’m thinking that you’ve got something here that is in its early days.
Eric Bassier: We view this as an area where we are going to grow. We actually think tape is going to be more relevant because of this type of a use case. But it’s important to put that in context. The overall tape storage market is still under $1 billion versus the disk market and the flash market, which are massive – many, many billions.
What I will say, though, is the tape business is growing. Our tape revenues are increasing. And the reason is because of the way that the largest hyperscalers are using tape. Effectively, they’re using it behind object stores with software that they’ve developed themselves. And here is the premise of our strategy.
The whole way we’ve developed the portfolio we have is our belief that HPC organizations have basically the same need at maybe a slightly smaller scale, and maybe a few years back, too. They’re not going to invest the four or five years of engineering time to develop their own object store stack code. So we’ve said: we’ve built this for you, we’ve put it in a box. If you’d like AWS Glacier, you don’t want to put all your HPC data in the public cloud, we’ve built Glacier in a box for you. We can deploy it at your site, we can deploy it at multiple sites. We could, and where we’re going is, maybe we might host part of that for you. And that’s where our roadmap takes us.
Blocks & Files: So that could be as a Quantum cloud. You could have, for example, some Quantum ActiveScale systems in a colocation sites or an Equinix centre or something like that, and make that available to HPC customers?
That is something that we are considering. And just to make a comment on Genomics England. They take advantage of the capability of our ActiveScale object store software to do what we call geo-spreading. So they have object store systems that are deployed at three sites, and the ActiveScale software geo-spreads to the objects across all three sites. So you actually have a single system, a single namespace, that’s spread across three sites. And we can do that either on disk or tape. So conceptually, you can have disk at three locations and a tape system at three locations. But it’s a single namespace, a single object store.
But we have many customers that say, well, I might have two sites, but I don’t have three or, you know, I’ve got only got one site. Would Quantum be willing to host the other two? And that’s where I think our our customers are going to lead us there in terms of what’s the right model.
Blocks & Files: So you could provide a component of a customer’s private cloud?
Eric Bassier: Correct. I think really, you know, private cloud is one of these things that means different things to different people. But that is where we’re getting a lot of the early customer engagements that we have. That is how they’re expressing their initiative. They’re saying, ‘hey, we want to build a private cloud for archival data’. And we say, ‘we can help you build a private cloud for your archival data’. So yes, we think we’re pretty excited about that.
Comment Blocks & Files thinks Quantum is well positioned with StorNext and ActiveScale using both disk and tape to pick up a number of HPC customers as they accumulate too much data to store on their primary and possibly flash storage systems and tier older data off to nearline disk and then to tape. The single namespace and geo-spreading unifies tape and disk into effectively a single object store and that potentially makes life easier for HPC admin staff.
Researchers at nano- and digital technology R&D center IMEC are suggesting memory could become liquid in the future.
Engineering boffins at Belgium’s IMEC (Interuniversity Microelectronics Centre) presented a paper on liquid memory at the 2022 International Memory Workshop (IMW). Maarten Rosmeulen, program director of Storage Memory at IMEC, identifies DNA storage as a post-NAND high-density but slow archival technology in an IMEC article. He proposes that liquid memory could replace disks for nearline storage at some point in the future.
Two types of memory, colloidal and electrolithic, are said to have the potential for ultra high-density nearline storage applications and might have a role between disk and tape from 2030 onwards “at significantly higher bit per volume but slower than 3D-NAND-Flash.” Rosmeulen says: “We anticipate that with these approaches, the bit storage density can be pushed towards the 1Tbit/mm2 range at a lower process cost per mm2 compared to 3D-NAND-Flash.”
In colloidal memory, two types of nanoparticles are dissolved in water contained in a reservoir. They carry data symbols. The reservoir has an array of capillaries through which the nanoparticles flow one at a time. “Provided that the nanoparticles are only slightly smaller than the diameter of the capillaries, the sequence in which the particles (the bits) are entered into the capillaries can be preserved.” The bit sequences encode information and the nanoparticles can be sensed by CMOS peripheral circuit-controlled electrodes at the entrance to each capillary tube for writing and reading.
Frequency-dependent dielectrophoresis is being investigated as a write mechanism. “A selective writing process can be created by choosing two particles that respond differently to the applied frequency (attractive versus repulsive).” R&D with polystyrene nanoparticles is ongoing “to fine tune the concept and provide the first proof of principle on a nanometer scale.”
Electrolithic memory also has a fluid reservoir with an array of capillary tubes. Two kinds of metal ions – A and B – are dissolved in it and electro-deposition and electro-dissolution techniques are used for reading and writing information.
There is a working electrode at the base of each capillary tube, which is made of an inert metal like ruthenium, and the reservoir also has a single counter electrode. A CMOS integrated circuit connects to the dense array of working electrodes. The common counter electrode plus the reservoir and a working electrode form an electrochemical cell for each capillary tube.
The articles says: “By applying a certain potential at the working electrode within the capillary, thin layers of metal A can be deposited on the electrode. Metal B will behave similarly but deposits at a different onset potential – determined by its chemical nature.” Information can be encoded in a stack of alternating layers on each working electrode. For example, “1nm of metal A can be used to encode binary 0, while 2nm thick layers of A encodes a binary 1. A layer of metal B of fixed thickness (e.g. 0.5nm) can be used to delineate subsequent layers of A.”
Rosmeulen says: “These new liquid-based memories are still in an exploratory research stage, with the electrolithic memory being the most advanced. Nevertheless, industry has already shown considerable interest in these concepts.”
HJe adds: “To become a viable storage solution for nearline applications, the technology must also have adequate response time, bandwidth (e.g. 20Gbit/s), cycling endurance (103 write/read cycles), energy consumption (a few pJ to write a bit), and retention (over 10 years). These evaluations will be the subject of further research, building on IMEC’s 300mm liquid memory test platforms with both colloidal and electrolithic cells in different configurations.”
Dell grew Q1 2023 revenues to a record level on a year-over-year basis with commercial PCs up 22 percent, server revenues up 16 percent, and storage up 9 percent, while still grappling with a troublesome supply chain.
In the quarter ended April 30, Dell revenues were $26.1 billion, up 16 percent: the Infrastructure Solutions Group – servers, networking and storage – delivered $9.3 billion, and the Client Solutions Group – PCs – contributed $15.6 billion. Profit $1.1 billion profit, a rise of 17.3 percent on a year ago. This was Dell’s fifth consecutive quarter of growth.
Vice chairman and co-COO Jeff Clarke said in a statement: “We followed a record FY22 with a record first quarter FY23… with growth across our business units.”
Co-COO Chuck Whitten talked of the benefits of having “a strong, geographically and sector-diverse business covering the edge to the core data center to the cloud,” proclaiming: “We are positioned to pursue growth wherever it materializes in the IT market, given the predictability, durability and flexibility in our business.”
Financial summary
Operating income: $1.6 billion, up 57 percent year-on-year
Operating cash flow: -$0.3 billion, primarily driven by annual bonus payout and seasonal revenue decline
Diluted EPS: $1.84, up 36 percent and a record
Recurring revenue: $5.3 billion, up 15 percent
Cash & Investments: $8.5 billion, up 15 percent
Core debt, which was at $42.7 billion in fiscal 2019, has fallen to $16.5 billion as of Q1 2023. That’s down 41 percent year-on-year.
Within ISG, servers and networking provided $5 billion in turnover, up 22 percent, with storage up 9 percent and generating $4.2 billion. This follows several quarters of slower storage sales.
Analyst Aaron Rakers told subscribers: “Dell delivered what we think should be considered very strong F1Q23 ISG results.”
CSG saw commercial PC revenues rise 22 percent to $12 billion and consumer revenues increase 3 percent to $3.6 billion.
Rakers said: “Dell’s strong CSG (PC) results reflect the company’s ability to continue to leverage a leadership position in software + peripherals – something we expect to be continually driven by the need for enterprises to support hybrid work.”
ISG is a key pillar of Dell’s business and the company believes it has ample room to grow. The ISG total addressable market was $165 billion in 2021, and is forecast to grow to $216 billion in 2025. Dell noted a shift in IT spend during Q1 from consumers and PCs to datacenter infrastructure, and expects CSG growth to moderate during fiscal 2023. It will moderate most in consumer PCs and ones running Chrome, less so in commercial PCs, which represents over 70 percent of Dell’s PC business.
Earnings call
In the earnings call, Clarke said: “We continue to execute quite well in a complex macro environment… We experienced a wide range of semiconductor shortages that impacted CSG and ISG in Q1. In addition, the COVID lockdowns in China caused temporary supply chain interruptions in the quarter. As a result, backlog levels were elevated across CSG and ISG exiting the quarter.
“We expect backlog to remain elevated through at least Q2 due to current demand and industry-wide supply chain challenges.” But the supply chain problems are being worked through. “People are coming to Dell because there’s confidence in our supply chain to deliver.”
Whitten said IT budgets remain healthy. “What we don’t see is an immediate move to go after a reduction in IT budgets. I mean, right now, it is a very healthy infrastructure environment.”
Revenues in the next quarter are expected to be between $26.1 billion and $27.1 billion, up 10 percent at the mid-point. Full fiscal 2023 revenues are being guided to grow approximately 6 percent over 22’s $101.2 billion. This is despite macroeconomic concerns such as the geopolitical environment, inflation, interest rates, slowing economic growth, currency, and continued disruption to supply chains and business activity.
Dell expects its relationship with VMware to continue unchanged if Broadcom completes its purchase.
Object storage software supplier Scality says its S3 Object Storage is available on the HPE GreenLake Cloud Services platform for on-premises use with cloud-like data services that meet data sovereignty requirements. HPE and Scality have co-deployed over an exabyte of storage, with hundreds of joint customers in more than 40 countries. The two say they provide freedom from being beholden to Amazon or the public cloud – therefore, no expensive data access or egress fees. Read up on the background here.
…
Amazon has added new file services to AWS. AWS Backup now allows you to protect FSx for NetApp ONTAP. Amazon EFS has increased the maximum number of file locks per NFS mount, enabling customers to use the service for a broader set of workloads that leverage high volumes of simultaneous locks, including message brokers and distributed analytics applications. Amazon FSx for NetApp ONTAP is now SAP-certified for workloads including S/4HANA, Business Suite on HANA, BW/4HANA, Business Warehouse on HANA, and Data Mart Solutions on HANA. A Jeff Barr blog provides more details.
…
Data protector N-able announced N-hanced Services for its MSP partners with onboarding, support, custom solutions, and migration services to assist MSPs with integrating and consolidating remote monitoring and management platforms. Leo Sanchez, VP of support and services at N-able, said: “We know that our partners are struggling with challenges like labor shortages, security, and more and more devices in more increasingly complex environments. With this new offering, we’re putting our experts to work to help them get the most out of our solutions and optimize their current workforce so they can actually do more with less staff. They can leverage bespoke services that meet them where they’re at and get to a place where they see real value faster.”
…
NetApp has closed the acquisition of Instaclustr, a database-as-a-service startup. A blog by EVP Anthony Lye said: “We just closed our acquisition of Instaclustr, marking a huge step in the transformation of NetApp. Think about it, in seven years, an on-premises storage company has built unique relationships with the three biggest clouds. Today, we’re offering a rich set of data services and hold a leading position in Cloud Operations (CloudOps).” He claimed: “Instaclustr allows us to deliver on our promise of more cloud at less cost.”
…
Nutanix has signed up the UK’s BUPA private health organization as a customer, providing a two-site IT platform (100 nodes in total) for BUPA’s 3,000-plus Citrix desktop-as-a-service (DaaS) users. BUPA has also begun using Nutanix Calm to automate management of its DaaS computing system and allow for deployment of this and other workloads to any public cloud. BUPA said: “Citrix modules that previously took half an hour or more to boot were now starting in seconds.” Read a case study here.
…
High-performance NVMe array supplier Pavilion Dataannounced a partnership with Los Alamos National Laboratories (LANL) to co-develop and evaluate the acceleration of analytics by offloading analytics functions from storage servers to the storage array, minimizing data movement by enabling data reduction near the storage. LANL is moving their I/O from file based to record or column based, to enable analytics using tools from the big data community. It has shown 1000x speed-ups on analytics functions by using data reduction near the storage devices via their DeltaFS technology. We are told the data processing algorithms of Pavilion HyperOS coupled with the performance density of its HyperParallel Flash Array provides a fast computational storage array capability enabling analytics offloads at scale.
…
Qumulo has introduced non-disruptive rolling Core software upgrades across a cluster of nodes. It’s available immediately for Network File Storage (NFS) v3.x, with a planned expansion to support NFS v4.x this summer. The rolling update process within Qumulo Core allows for node reboots with no downtime to the cluster. Stateful requests can be handed off to other processes on an individual node, as other processes update. Full platform upgrades update underlying component firmware and operating systems, giving the administrator an option to perform the upgrade as a parallel upgrade (with minimal downtime) or as a rolling, non-disruptive upgrade (with one node going offline and upgraded at a time). Non-disruptive rolling upgrades give storage administrators the freedom to run upgrades during normal operating hours instead of during off-hours, weekends or holidays.
…
Data protector and security supplier Rubrik has hired Michael Mestrovich as CISO. At the Central Intelligence Agency, Mestrovich led cyber defense operations, developing and implementing cybersecurity regulations and standards, and directing the evaluation and engineering of cyber technologies. He also served as the Principal Deputy Chief Information Officer for the US Department of State where he was responsible for managing the department’s $2.6 billion, 2,500-person global IT enterprise.
…
Snowflake’s revenue for its first quarter of fiscal 2023 ended April 30 was $422.4 million, up 85 percent year-on-year. Product revenues were up 84 percent to $394.4 million. The net revenue retention rate was 174 percent. It now has 6,322 total customers, up from 5,994 last quarter, and 206 customers with trailing 12-month product revenue greater than $1 million. Snowflake recorded a loss of $165.8 million. Chairman and CEO Frank Slootman said: “We closed the quarter with a record $181 million of non-GAAP adjusted free cash flow, pairing high growth with improving unit economics and operational efficiency. Snowflake’s strategic focus is to enable every single workload type that needs access to data.” Guidance for the next quarter is for revenues of $435-$440 million, compared to $272.2 million revenues in Q2 2022.
Snowflake’s revenue growth is accelerating
…
NAND industry research house TrendForce says NAND flash bit shipments and average selling prices fell by 0.5 percent and 2.3 percent respectively in Q1 2022, causing a 3 percent quarterly decrease in overall industry revenue to $17.92 billion. The market was oversupplied, resulting in a drop in contract prices in Q1, among which the decline in consumer-grade products was more pronounced. Although enterprise SSD purchase order volume has grown, demand for smartphone parts has weakened and inflation is rising. Looking to Q2, the same dynamics are expected to continue to slow the growth of consumption, however, the ongoing shift large North American datacenters to high-capacity SSDs will drive enterprise SSD growth by 13 percent.
…
Virtium has announced a new StorFly XE class M.2 NVMe SSD product portfolio with up to 10x more endurance than TLC SSDs – because they use pseudoSLC flash, a hybrid of 2-bit per cell MLC using firmware to emulate the storage states of 1-bit per cell SLC, with up to 30,000 PE cycles. This pSLC NAND has much better data retention performance at higher temperatures, yet at a fraction of the cost compared to SLC. The XE class M.2 NVMe SSDs are configurable and include vendor-specific commands to tune critical parameters including power and capacity. They also provide extremely steady performance over the full SSD capacity and over the full -40 to 85°C temperature range, meaning they won’t suffer the frequent and erratic performance drops often found in client and enterprise-class SSDs.
Japanese disk drive media manufacturer Showa Denko K K (SDK) is shipping 26TB capacity disk drive platters. It has not revealed its customer, but Toshiba uses Showa Denko platters in its nearline 3.5-inch disk drives.
Nearline disks are high-capacity (15TB-plus) drives spinning at 7.200rpm, typically with a SATA interface.
SDK says its new disk media supports energy-assisted magnetic recording and shingled magnetic recording (SMR) because its has fine crystals of magnetic substance on the aluminum platter surface. It includes technology to improve the media’s rewrite-cycle endurance. SDK says it produces disk platter media by growing epitaxial crystals at the atomic level, forming more than ten layers of ultra-thin films with a total film thickness of no more than 0.1μm.
SMR drives need host management of data writes to zones of overlapping write tracks – a data change in any part of the zone requires the whole zone to be copied, edited, and rewritten. This effectively limits their use to hyperscaler customers who can have their system software modified to manage SMR drives. Such drives are unlikely to appear in standard enterprise datacenters or video-surveillance farms or even desktop PCs, unless system software is altered to support them.
Like Western Digital, Toshiba uses energy-assisted recording to strengthen data writing in the very small bit areas of 18TB-plus drives. Its specific technology is called MAMR (microwave-assisted magnetic recording). Seagate has an alternate HAMR (heat-assisted magnetic recording) technology in development.
Western Digital and Seagate are shipping 20TB conventionally recorded (CMR) nearline drives with Toshiba at the 18TB level. Seagate is sample shipping a 20TB-plus drive with Western Digital sample shipping 22TB CMR drives and 26TB shingled magnetic recording (SMR) drives. SDK and Seagate are partnering to build HAMR disk drives.
Toshiba has a 20TB CMR HDD on its roadmap and a 26TB SMR drive. The Showa Denko 26TB HDD media news indicates that this could be announced for sample shipping quite soon.
We could see both Western Digital and Toshiba shipping 26TB SMR drives in the summer but not Seagate, unless its 20TB-plus drive has an SMR variant at the 26TB capacity level.
SDK said it plans to produce nearline disk drive media with a greater than 30TB capacity by the end of 2023.
Hyperconverged infrastructure software vendor Nutanix is warning that supply chain issues and sales reps leaving for pastures new will result in weaker-than-expected sales for the remainder of its financial year.
Revenues in the company’s Q3 of fiscal 2020 ended April 30 were $404 million, up 17 percent year-on-year, beating the company’s own forecasts of between $395 million to $400 million. It reported a net loss of $111.6 million, versus a net loss of $123.6 million in the year ago period.
President and CEO Rajiv Ramaswami said during an earnings call: “Our third quarter reflected continued solid execution, demonstrating strong year-over-year top and bottom line improvement.
“Late in the third quarter, we saw an unexpected impact from challenges that limited our upside in the quarter and affected our outlook for the fourth quarter. Increased supply chain delays with our hardware partners account for the significant majority of the impact to our outlook, and higher-than-expected sales rep attrition in the third quarter was also a factor. We don’t believe these challenges reflect any change in demand for our hybrid multicloud platform, and we remain focused on mitigating the impact of these issues and continuing to execute on the opportunity in front of us.”
Full FY22 revenues are now forecast to be $1.55 billion, an 11.2 percent year-on-year increase. Nutanix had predicted a revenue range of $1.625 billion to $1.63 billion three months ago.
Significant net loss improvement in the two most recent quarters
Customers run Nutanix software on third-party server hardware and shipping delays impacted revenues and orders. Rukmini said: “We saw these supply chain challenges impact us late in Q3, which limited our upside in Q3, and we expect these trends to continue in Q4.”
It won’t be just a short-lived issue, Ravaswami said: “We expect that these challenges in the supply chain are likely to persist for multiple quarters.“ He added that Nutanix wasn’t seeing any change in underlying demand.
Annual Recurring Revenue of $1.1 billion was up 46 percent on the year.
CFO Rukmin Sivaraman, said: “Q3 sales productivity was in line with our expectations… We saw rep attrition worse than in Q3, resulting in lower-than-expected rep headcount entering Q4. Under the leadership of our new Chief Revenue Officer, Dom Delfino, our sales leaders remain focused on getting rep headcount to our target level via both better retention and increased hiring efforts.”
Nutanix’s prior CRO, Chris Kaddaras, left in October 2021 to join startup Transmit Security, just seven months after starting at Nutanix. Delfino was hired from Pure, commencing his role in December 2021. Five months in and he’s losing sales reps. They are going, Ramaswami said, to startups with “the promise of quick IPO riches.”
Ramaswami suggested things would start to improve after the next quarter. Delfino is working on increasing sales rep productivity, “doubling down on training and enablement,” as well as “improved territory coverage and [a] higher level of quota attainment.”
Wells Fargo analyst Aaron Rakers told subscribers that sales reps had left for other start-up opportunities because they were “unable to make quotas,” and said: “We believe this could take a few quarters to rectify.”
The customer count rose by 586 in the quarter to 21,980. This is down on the 660 added a year ago and the 700 in the prior quarter. Nutanix says it saw a year-over-year improvement in win rates against VMware and other competitors.
Free cash flow was minus $20.1 million. Ramaswami said Nutanix was working to change this: “We continue to prioritize working towards sustainable free cash flow generation in FY ’23.”
The fourth-quarter outlook is for a growth of between $340-$360 million, weighed down by customer order delays due to server hardware availability and the lower-than-expected sales rep headcount. At the $350 million mid-point, this is a 10.4 percent decrease on the year-ago Q4. William Blair analyst Jason Ader told subscribers this is “an eye-popping $89 million below consensus.”
Ramaswami said: “We don’t believe our reduced outlook is a reflection of any change in our market opportunity or demand for our solutions. That said, we are focused on mitigating the impact of these challenges and continuing to drive towards profitable growth. “
Zadara has surpassed 400 global points of presence spanning 24 countries, and said it plans to have more, with edge cloud locations near every major city in the world powered by its partner network.
It has an ecosystem of more than 250 managed service providers (MSPs), with 441 locations in every part of the globe. Partners include Seagate, Cyxtera, Sandz, Viatel, Lumen, NTT, KT Corporation, KDDI, Africa Data Centres, and Rakuten.
Zadara’s edge locations
Nelson Nahum, CEO and co-founder, said: “Zadara is realizing its vision to deploy and support a network of service providers that scale without the need to build new data centers, while offering cloud services anywhere in the world. Bringing a better than public cloud experience, with optimized price, performance, security and compliance to our partners, means faster times to innovation and new growth opportunities.”
Zadara’s view of its services
Nahum claimed Zadara has “more global points of presence than any other edge cloud provider.” But it is seemingly lagging its own targets as a July 2020 blog written by Nahum said: “By the end of 2021 we expect to have over 1,000 locations in our partner network allowing us to serve our customers with edge-like capabilities to solve for low-latency.” Even though it’s fallen short, it claims it has more edge points of presence than other public clouds.
The company provides compute (EC2-compatible), storage (block, file, object), and networking facilities as managed services located on-premises, in hybrid and multi-cloud, and at the edge. Customers can have the hardware in their central datacenters or edge locations and lease it from Zadara, and also use Zadara’s IT resources in the main public clouds and its own public cloud. This is an alternative to AWS, Azure, Google, and Oracle’s cloud.
Zadara’s table comparing its own cloud to AWS, Azure, and Google
Zadara has doubled its headcount during the past 12 months and expanded its corporate presence in Tokyo, Tel Aviv, and throughout the US, where it is headquartered. The company started out from its Israeli base by offering managed storage-as-a-service, and has since expanded into compute, by acquisition, and networking (infrastructure-as-a-service).
A blog says: “After 17 consecutive quarters of growth, we are proud to share our gross profit year-over-year has increased 93 percent This was driven by the strength of our partner network and 100 percent increase in new customers year-over-year.”
Where’s the money for this growth coming from? The company has raised more than $60 million in six rounds since it was founded in 2011, with the last being an undisclosed venture round in 2019. This followed a $25 million raise in 2018. We think there is a substantial element of self-funding from revenues here.
It’s recently entered Formula 1 Grand Prix sponsorship, having a relationship with Sauber Motorsport, which runs the Alfa Romeo F1 Team ORLEN. This suggests it has marketing money to burn.
All the major system and storage suppliers now offer subscriptions, such as HPE GreenLake and Dell APEX, with deals for MSPs. Presumably Zadara has a competitive offer concerning costs, performance, points of presence, support, etc, yet it has to source its own hardware and software from original equipment manufacturers. You would think this would give the OEMs a pricing edge yet Zadara is growing and growing. Should we view this as being like some airlines leasing airplanes from intermediaries who buy them from the manufacturers?
We look forward to Zadara filing for an IPO because then we should get to see details of how its business operates and its financial performance. Until then we’ll watch it grow some more and maybe get acquired.
A private-equity group that has put money into SoftIron claims the integrated hardware-software array manufacturer has the lowest carbon footprint of storage servers when compared to Dell, HPE, and Supermicro.
Environmental, Social and Governance (ESG) reporting is becoming widespread in North America and Europe. IT organisations are looking to their suppliers to help them lower their carbon footprint.
UK-based investing group Earth Capital focuses on the development and deployment of sustainable technology, and is SoftIron’s sole VC backer. It has issued a SoftIron Carbon Footprint Report 2022 as part of a Net Zero Asset Managers Initiative committing it to support the goal of net zero greenhouse gas emissions by 2050 or sooner. Earth Capital has conducted net-zero assessments on assets under its management, including SoftIron storage systems.
Earth Capital said it found “SoftIron’s product has a greatly reduced carbon footprint mainly on account of the comparably low energy usage during operation.”
It claimed the power required to store 1TB of data with SoftIron’s HyperDrive system was between 2 and 4.6 times less than that needed for similar Dell, HPE, and Supermicro systems.
“The SoftIron appliances produce up to 80 percent less heat than the standard, thereby radically reducing emissions pertaining to cooling of data centers,” the report says.
The researchers calculated the power needed to store 1TB of data as a proxy for a device’s carbon footprint. They took into account the product’s lifetime, operating power consumption, and assembly and use locations. Their assessment used publicly available data to compare the power required to store 1 TB of data using a SoftIron appliance to the power required using alternative Ceph reference architectures.
The report states: “The full life cycle emissions of a single SoftIron node equates to 2,898 ± 290 kg CO2e. Of these emissions, ~77 percent (2227 ± 223 kg CO2e) pertain to the use of the product while ~23 percent (671± 40 kg CO2e) pertain to the collective contributions of raw materials, manufacturing, transportation, and installation (fig.1). These findings highlight that electricity during product operation is a significant carbon source for SoftIron. However, carbon emission pertaining to energy usage in industry standard servers is ~11,135 kg CO2e (i.e. up to 5 times larger than that of SoftIron.)
A chart shows the results of Earth Capital’s comparisons with Dell PowerEdge R740xd systems, HPE ProLiant DL380 and Apollo 4510 products, and Supermicro 6019P and 6049P SuperStorage systems with SoftIron’s HyperDrive HD11120:
Supermicro’s 6049P was second best with 2.8W/TB needed to store 1TB of data compared to SoftIron’s 1.2W/TB. Dell was third best with 2.9W/TB while HPE’s Apollo was rated at 4.1W/TB. The ProLiant needed 5.1W/TB, while Supermicro’s 6019P was the worst system at 5.5W/TB.
The energy consumption comparisons do not account for additional carbon savings stemming from reduced heat generation.
Dell, HPE and Supermicro have been contacted for comment but not replied at publication time. Subsequently Supermicro stated: “it will pass on responding this time around. Supermicro has a continued interest in the carbon footprint of its own and customers operations, and will publish a green data centre report this summer.”
The Earth Capital report notes: “Many market-leading data centre suppliers acquire commoditised, off-the-shelf hardware which use the latest iteration of central processing units (CPUs). Research indicates that each iteration of CPU requires more power than its predecessor (fig.3). Therefore, the industry standard data centre appliances are in fact becoming less energy-efficient, thereby creating more financial and environmental costs for data centres.”
Earth Capital’s fig 3 chart looks like this:
The report concludes: “Many current customers are seeing SoftIron’s energy efficient approach as a steppingstone towards their sustainability goals. For example, a recent customer (THG) has adopted SoftIron’s solutions as part of a key corporate goal to maintain CarbonNeutral certification as they grow thier footprint of data centres.”
There is no industry-standard way to measure a storage (or other) system’s carbon footprint. The SNIA, with its existing Green Storage Technical Workgroup and Green Storage Initiative, could perhaps have a look at developing one.
Favorable economics – data-intensive applications and workflows fuel new tape growth due to its significant total cost of ownership (TCO) advantages
Security – tape’s inherent air gap provides additional levels of cybercrime defense
Data accessibility – tape performance improves access times and throughput
Sustainability – tape plays a significant role in green data center strategies
Optimization – tape-based active archives boost storage optimization, providing dynamic optimization and fast data access for archival storage systems
Let’s unpack that. The favorable economics basically means tape storage costs less than disk storage on a $/TB basis, and both are less than flash storage. However, tape is slow, with a longer time to first byte than disk or NAND, and this makes it the choice for rarely accessed data where latency is less of a concern than storage cost.
Tape’s inherent physical air gap is a good point but the virtual air gaps marketed by backup storage, unstructured data storage, cloud storage, and cybersecurity companies are generally acknowledged to be effective and so its superiority here is lessened.
The data accessibility point is harder to understand. Sure, tape is better than disk in bandwidth terms. As the report says, “HDDs and SSDs have faster access times to the first byte of data. For large files, tape systems have faster access times to the last byte of data… The LTO-9 and TS1160 enterprise [tape] drives each have a data transfer rate of 400 MB/sec. This compares to the 7,200 rpm HDDs ranging between 160-260 MB/sec.”
But typically a large file would be striped across disk drives in an external storage array and read back from several drives in parallel, nullifying the single tape drive’s advantage. Also many primary and several secondary data storage devices these days use SSDs and not disk drives. In bandwidth terms that’s pretty much game over, especially with NVMe SSDs. In our view this claimed advantage evaporates in front of our eyes when we look at it.
Yes, tape is getting faster, generation by generation, but it’s still slower than disk drives and all-flash arrays. The evangelists implicitly accept this as we shall see when we get to the fifth trend.
The council’s fourth point about sustainability is a good one. Obviously non-streaming, off-line tape cartridges stored on a shelf don’t need power and cooling. But this trend supports its use only if the need for fast data access has already been discounted. No one is going to use it for primary data storage if it means that your servers can only support tens of transactions a second instead of thousands, not even to save some thousands of tonnes of carbon emissions a year.
And so we get to the fifth point, the active archive. The diagram in the report shows a cache buffer placed in front of a tape library:
A diagram from the Tape Storage Council’s report
It is operated by a server running data management software, which presents a file or object interface upstream and a tape interface downstream to the library. The report says: “An active archive integrates two or more storage technologies (SSD, HDD, tape, and cloud storage) behind a file system providing a seamless means to manage archive data in a single virtualized storage pool.”
Archive storage in the cloud uses tape so we are really talking about disk/SSD front ends and a tape backend here. And why have this cache buffer? It “provides dynamic optimization and fast data access for archival storage systems.” The “SSDs or HDDs serve as a cache buffer for archival data stored on tape providing faster access to first byte of data, higher IOPs, and random access.”
In other words, you need a disk or SSD buffer because tape data access is slow. Ideally people wouldn’t use tape at all because it’s far too slow, but it is cheap and reliable so we put up with it and ameliorate its slowness with a cache buffer in active archive setups.
The report says the need for data archives will grow due to increased cold data storage needed by cloud, HPC, IoT, life sciences, media and entertainment, video surveillance, and sports video. That’s most likely true but it’s still standard data archiving on tape. It won’t play a broader role. It will play the same role it always has.
Tape is not dead yet and its capacity-increase roadmap is impressive. It’s actually thriving. But that is not because of its wonderful technology, good though it is, but because there is nothing better. It’s cheap, it holds an awful lot of data, it’s reliable, it’s slow, and it works. Enough said.
