LogoWatch Today Toshiba Memory Holdings officially rebrands as KIOXIA and has unveiled the new logo.
It’s… grey.
What do the marketing folks say? “The silver of Kioxia’s new logo will be the company’s official corporate colour, meant to represent the superior quality of its memory technology. In addition to silver, the company will have communication colours including, light blue, magenta, light green, orange, yellow, white and black.”
Kioxia is pronounced ‘KeeOxchia’ and should be printed in all-caps.
Naohisa Sano, Kioxia’s chief marketing officer, said: “Our full brand colour palette of bright, vibrant colors represents KIOXIA’s fun, future-driven culture, and passion for using memory to create new experiences and a colourful future for the world. Our new corporate logo and brand identity better reflect KIOXIA’s mission and vision to uplift the world with memory, using technological innovation to create new value for society.”
Kioxia’s logo is black on its website.
Consumer retail products such as solid state drives, SD cards and USB flash memory, will be released under the Toshiba Memory brand name until the end of the year. Kioxia kicks in in January 2020.
And vulnerable. The ongoing trade spat between the Trump administration and China has seen restrictions imposed by the US Department of Commerce on Huawei that bar US firms trading with the Chinese tech giant. The US delayed the trade ban for 90 days on August 2019.
China’s plans
The two biggest types of semiconductors are DRAM and NAND.
The DRAM market has three suppliers: Samsung, SK Hynix and Micron. The combination of new fab cost to build the latest DRAM technology, lowish market demand growth and areal density increase, means that the industry cannot afford the entry of another supplier. If China Inc. enters the DRAM market then it will be a fourth supplier. Handy’s assessment is that the DRAM industry will contract back to three suppliers. Either Samsung, SK Hynix or Micron will exit the industry or be acquired.
Example chart from report.
Blocks & Files suggestion is that Micron will buy SK Hynix and so add DRAM and NAND capabilities in a single move.
The NAND industry has higher demand growth and its suppliers – Intel, Micron, Samsung, SK Hynix, Toshiba and Western Digital – can withstand the entry of China Inc. Handy thinks China’s YMTC will enter mass NAND wafer production in 2023. This will extend the current over-supply for a year or two and depress vendor earnings.
The report is detailed and covers the effect of China’s entry into the DRAM and NAND markets on tool suppliers and OEM customers. Handy suggests Western suppliers will need to partner with Chinese companies to progress in the Chinese market.
Dell EMC has announced a mid-range ECS object store plus additional security software features to suit US DoD needs.
Effectively it has added higher-capacity drives to the entry-level chassis to create a bigger system that should cost much less than a high-end box.
Dell EMC positions ECS as building on the legacy of Centera and Atmos object storage systems. The ECS500 is a fourth generation ECS appliance and fits mid-way between the gen 3 ECS300 and ECS3000 appliances.
EX300 – 12 disks/2U node – 1TB, 2TB,4TB or 8TB drives – to 1.54PB/rack
EX500 – 12 or 24 drives/2U node – 8TB or 12TB drives – to 4.6PB/rack
EX3000 – 12TB drives – 45 or 90 disks/4U node – to 8.6PB/rack
Comparisons
For comparison, 1U and 4U Cloudian HyperStore appliances can use 14TB drives with up to 12 or 70 per chassis. The maximum capacities per chassis is 980TB. These Cloudian systems also store metadata in SSDS for faster access.
NetApp StorageGRID systems come in 2U x 12 drive, 4U x 60 drive and 5U x 58 drive chassis. With 12TB drives the maximum chassis capacity is 720TB.
The ECS range
Back in the Dell EMC camp the EX300 has 10GbitE front and backend access whereas the EX500 and EX3000 use faster 25GbitE. There is one server per node with the EX300 and EX500 systems and the EX3000 can have one or two servers per node.
The 2U and 4U ECS chassis
The EX3000S has a single server and there are a minimum of five nodes and a maximum of eight per rack. There are two servers in the EX3000D chassis and between six and 16 in a rack. Like the EX3000S, the EX300 and EX500 have a minimum of five nodes.
The EX3000S has either 45, 60 or 90 drives while the EX3000D supports 30 or 45 drives: that extra server takes up space in the chassis. Dell EMC said all nodes within an EX3000 rack must always have the same hard drive configuration.
EX300 nodes of differing drives sizes can be mixed if they are added in groups of four nodes at a time.
There are up to 16 EX500 appliances (nodes) per rack. All ECS systems scale out beyond a single rack.
There are several news items about Intel Optane in this week’s roundup, as Intel’s Optane software ecosystem building efforts pick up. But Cloudera, first.
Cloudera’s cloud data warehouse
Cloudera is now competing with Snowflake and Yellowbrick Data. The company has launched the Cloudera Data Warehouse, a cloud-native service for self-service enterprise analytics on the new Cloudera Data Platform (CDP). It enables quick deployment and easy administration of cloud data warehousing, moving on-premise workloads to the cloud with consistent security and governance. It ingests data anywhere, at massive scale, from structured, unstructured, and edge sources.
Anupam Singh, GM of data warehouse at Cloudera, said: “Hundreds of users can simply provision their own resources at the click of a button and analyse all data together, wherever it is on-premises or in the cloud, without breaking the bank, without breaking metadata and security, and lastly, without locking data into proprietary formats and silos.”
Dell Analysts’ day
William Blair analyst Jason Ader attended a Dell Analysts’ Day last week and sent subscribers this summary;
Dell believes that given increasing IT complexity, customers will look to consolidated strategic vendors, with Dell uniquely positioned to serve that role;
Dell’s focused execution in the storage market could allow the company to reclaim lost market share and increase competitive pressure on pure-play vendors like NetApp;
Continued integration of VMware’s portfolio positions Dell to be a leader in private/hybrid cloud infrastructure, though Dell will continue to support VMware’s open structure, allowing it to partner and integrate with many of Dell’s competitors;
Dell is betting big on AI/ML technologies and integrating them into its products, leveraging Dell Technologies Capital to gain exposure to emerging companies (e.g. it has invested in Noodle.ai and Graphcore);
Dell remains focused on improving profitability and deleveraging (core debt remains at $36.4bn; leverage ratio of 3.6 times), which will limit Dell’s ability to use spending to spur growth and/or its balance sheet for major M&A activity.
Ader points out Dell’s storage business has reclaimed 375 basis points of previously lost share. Dell believes that it is well positioned to continue reclaiming share, targeting a reclaim of roughly two-thirds of previously lost share as feasible.
Dell has made structural changes, including cutting down the number of products from 88 to 22, increasing sales-and-marketing spend, and reorganising R&D to be more collaborative and focus on newer products.
It has integrated VxRail with vSAN and put them on the same release schedule, which has simplified the company’s HCI offering and messaging. The stronger integration of Dell EMC and VMware could put more pressure on Nutanix.
Ader thinks that room remains for both Dell/VMware and Nutanix to be successful (currently a two-horse race competitively).
Hazelcast in-memory software and Optane
In-memory software business Hazelcast has agreed with Intel to use Optane for real-time applications, artificial intelligence and internet of things systems for enterprises. Its Project Veyron is focused on running Hazelcast technologies on Intel second generation Xeon Scalable processors and Optane DC Persistent Memory (DIMMs). Hazelcast said Project Veyron will accelerate the completion of parallel in-memory tasks, complex analyses for more sophisticated models and the use of structured and unstructured data sets.
Raker’s take on IDC HCI report
Wells Fargo senior analyst Aaron Rakers has added more intelligence to the bare bones IDC storage tracker HCI report, published last week. He told subscribers that IDC estimates that the total HCI market revenue at $1.825bn, saw continued deceleration of y/y growth in 2Q19 at +24 per cent y/y, compared to +69 per cent, +57 per cent, and +47 per cent y/y in 3Q18, 4Q18, and 1Q19.
Is HCI growth tailing off?
Rakers thinks the most notable takeaway is VMware’s continued strength with total vSAN + Ready Nodes revenue totaling $694M in C2Q19, +39 per cent y/y and equating to a ~38 per cent market share.
This compares to Nutanix‘s software + hardware revenue (including third party OEMs) which totaled ~$522M in C2Q19, +5 per cent y/y and implying a ~29 per cent share (vs. 34 per cent a year ago). NetApp’s HCI revenue was estimated at ~$39M, up 113 per cent y/y, but was down 15 per cent q/q.
Hmm. NetApp isn’t exactly experiencing soaring growth with its Elements HCI product.
HP SimpliVityrevenue grew 30 per cent y/y to $125M, while Cisco’s HyperFlexrevenue stood at ~$114M, up 47 per cent y/y.
Intel has started shipping Stratix 10 DX field programmable gate arrays (FPGA). They support Intel Ultra Path Interconnect (UPI), PCI-Express (PCIe) Gen 4 x16 and a new controller for Intel Optane technology to provide high-performance acceleration.
Stratix supports select Intel Optane DC persistent memory dual in-line memory modules (DIMMs). They increase bandwidth and provide coherent memory expansion and hardware acceleration for upcoming select Xeon SP CPUs.
Intel Stratix FPGA.
The Stratix memory controller supports up to eight Optane DC persistent memory modules – Intel doesn’t like using the DIMM word – per FPGA (up to 4TB of non-volatile memory).
Intel said the UPI interface in combination with future select Xeon SPs should deliver 37 per cent lower latency (than Xeons without UPI) and improve overall system performance via coherent data movement and a theoretical peak transfer rate of 28 GB/sec.
The PCI Gen 4 x 16 interface delivers peak data bandwidth of 32 GB/sec. Apps should realise about two times more throughput.
VMware is one of many early access participants, and aims to develop coherent FPGA and CPU acceleration solutions.
Intel and penta level cell flash
At its Memory and Storage day in Seoul, Sep 26. Intel talked about penta level cell flash (5 bit/s cell). This is its slide.
The slide above shows PLC flash with 32 separate voltage levels that are detected by some read function in the chart on the right.
Intel is not committing to delivering a product but hints strongly at the possibility. A PLC SSD would have 25 per cent more capacity than QLC (4 bits/cell) flash and should have a lower cost/TB. With QLC flash having endurance around 1,000 write cycles, then PLC flash could be in the sub-500 range and be even more limited to read-centric workloads than QLC.
What Nebulon and its Nebunerds are building
In November 2018 Nebulon, a California startup founded by four 3PAR veterans, announced itself to the world as a ‘cloud-defined storage startup’. We covered the birth announcement at the time but no other details were forthcoming.
Since then we have gleaned some information about the company via Twitter and Linkedin posts made by company execs and also from recruitment ads.
The company has secured $14.5m from unnamed sources and employs at least 85 people – who it calls Nebunerds. It aims to drastically simplify storage and to deliver secure, scalable and powerful data insights by building cloud-defined storage.
This is the company’s term for software-defined storage and boils down to building a cloud back-end enterprise-class store for IoT edge devices. The store will be a scalable, efficient, secure and highly available cloud data platform for data ingestion, transformation, storage and computation.
Nebulon’s software will provide real-time monitoring, analysis and automation of its operations and of the data inside it. It is cloud-native, built from micro-services, and will run on AWS in the first instance.
We don’t know if there will be Nebulon agent software in the edge devices or if they transmit data to the back end via the S3 protocol. We anticipate early product news in the second half of next year.
Packet, Formulus Black and Optane
Developers using Packet’s cloud platform will be able to test, validate, and optimise data-intensive and real-time application workloads on Forsa.
Made by Formulus Black, Forsa OS runs application software re-encoded into bit markers and in memory.
Formulus Black has devised algorithms that optimise I/O between Optane DC persistent memory memory and CPU. Based on initial testing, the net result includes decreased CPU usage, more TPS/IOPS and lower latency under maximum load conditions
Forsa can pool memory from multiple CPU sockets on a single system or across systems. For instance, on a four-socket server, memory on all four CPU sockets can be pooled together. It can stitch memory from CPU sockets on a second server, creating memory-based storage devices with 12-24TB of capacity (assuming 512GB x 24 Optane DC persistent memory modules per server).
Forsa’s BLINK feature enables users to backup and restore data and system settings from Optane DIMMs to local SSDs, or a network storage target.
Also Forsa can increase the size of persistent memory block devices that are at full capacity, create clones and snapshots, and manage persistent memory resources on multiple nodes from one management console.
Quest Software’s V12.4 of NetVault Backup adds support for SAP HANA and Nutanix’ AHV hypervisor and better support for Office 365 Exchange Online and OneDrive. HANA and AHV support is via plug-ins.
Office 365 users can backup and restore user mailboxes to any cloud, disk or tape system. -based storage. Restoration is quite granular, down to files and emails which can be placed in particular Outlook folders.Users can back up and restore data from OneDrive user and group files and folders. Office 365 Active Directory and SharePoint Online support should arrive by the end of the year.
Microsoft ranks first as the most strategic vendor for customers, according to a double-blind market survey conducted by Dell in the first half of this year, Microsoft’s portfolio of Azure, Office 365, Teams, and DevOps is resonating with customers who view it as a one-stop shop for many of the technologies that underpin their digital transformation.
In the Dell survey Microsoft garnered 20 per cent of votes for most strategic vendor, IBM took second place with 13 per cent of votes and Dell placed third with 10 per cent (8 per cent attributed to Dell-EMC and two per cent to VMware),William Blair analyst Jason Ader reported in a note to subscribers.
Tom Bish, an IBM Systems Storage Subject Matter expert, introduces a YouTube video about “On Premise Storage Options for an OpenShift, Kubernetes environment.” He describes various Block and File Storage solutions, the tools available to manage them along with back-up and resiliency considerations. You can also learn about how the Container Storage Interface provides a conduit between your clusters and the storage devices you’ve selected.
Micron’s Crucial unit has announced the X8 Portable SSD with 500GB and 1TB capacities, sequential read/write speeds up to 1050 MB/sec, and a 3-year limited warranty.
Serverless search and analytics company Rockset announced the capability to analyse raw events from Apache Kafka in real time. Rockset’s tech enables SQL on NoSQL data.
The National Library of Scotland is using Scality RING object storage software to preserve and protect collections. It’s digitising 120 miles of shelving. There will be one copy in each of two on-premises RINGs in their Edinburgh and Glasgow data centres and a third copy in AWS Glacier Deep Archive.
Supermicro’s high-performance 4-Way MP SuperServer is available as an Intel select solution for SAP HANA. It enables customers to use Intel Optane DC persistent memory for SAP HANA scale-out servers.
Toshiba Memory America said its KumoScale storage software for NVM Express over Fabrics (NVMe-oF) v3.11 now supports Graphite and Prometheus telemetry frameworks. These integrations are built on the KumoScale REST API architecture, with adapters added for each new framework.
WANdisco has launched LiveAnalytics, a technology that makes migrated and migrating data immediately and continuously available for analysis. LiveAnalytics works in tandem with WANdisco’s LiveMigrator, so business operations and analytics can continue as data shifts to the cloud.
Zerto has announced the availability of Zerto 7.5 with expanded functionality with Azure, including support for Azure Managed Disks, scale-sets and Azure VMware Solution (AVS); integrations with HPE StoreOnce Catalyst; certification and support for VMware vSphere APIs for I/O Filtering (VAIO); and advanced analytics for reporting, planning and customisation of disaster recovery and long-term data retention.
People
Paul Forte
Actifio has a new sales boss, Paul Forte, who is called its Chief Revenue Officer. He was CRO at Channel Advisor and an EVP for North American Sales at Monster before that. Actifio told Blocks & Files that Ranajit Nevatia, SVP & GM Global Sales for Actifio GO & cloud business development, continues in place reporting to CEO Ash Ashutosh.
SaaS backup startup Clumio has poached Chad Kenny from Pure Storage and appointed him as VP and chief technologist. Kenny was VP for Products and Solutions at Pure. His responsibilities at Clumio will focus on building the Clumio technology partner ecosystem, guiding the longer-term product roadmap and promoting the right way to leverage the cloud for data protection.
Jai Menon, a former Dell and IBM exec, has joined Fungible as its chief scientist “to accelerate customers’ transition to a data-centric infrastructure.”
NetApp CMO Jean English left in August and is now CMO at Palo Alto Networks. Jeff McCullough, NetApp’s VP for the Americas channel, has also left, according to reports.
Nutanix has promoted Cyril VanAgt to lead iChannel and OEM activities in the Europe, Middle East and Africa region.
Tintri by DDN has hired Anand Ghatnekar as VP of engineering and Mario Blandini as chief marketing officer and chief evangelist.
Lorenzo Flores has been appointed as Vice Chairman for Toshiba Memory Holdings, effective in November, 2019. TMH is going to become Kioxia Holdings on October 1. Flores was Xilinx’s CFO and an Intel exec some time before that.
The pricing of Optane memory, also called 3D XPoint, will cannibalise some DRAM sales and this could give Micron a big headache.
“The revenues of 3D XPoint memory are likely to be matched by similar-sized sales declines in the server DRAM market,” writes Objective Analysis’s Jim Handy in a July 2019 update to his 3D XPoint Report.
On the other hand, the “SSD market for 3D XPoint will remain relatively small, with the new memory serving more to displace DRAM memory rather than NAND flash SSDs or HDDs,” he adds.
Since Optane DIMMs only work with newer and more expensive Xeon SP processors they will help Intel sell more of them to the extent that customers buy into this scenario.
Intel will care a lot about this as processors are its main source of income. Intel does not make DRAM and so will not care if DRAM sales are affected by Optane DIMMs.
Why Optane is cheaper than DRAM
Intel positions Optane as less expensive and slower to access than DRAM but more expensive and faster to access than NAND SSDs. An Objective Analysis chart, plotting storage, memory and cache types against bandwidth and price per GB, shows this positioning:
Jim Handy’s memory-storage hierarchy chart
Handy writes: “3D XPoint Memory won’t fit into the memory/storage hierarchy… unless it is sold for a lower price than DRAM. If its price is higher than DRAM’s price it will not be adopted.” It has, in effect a price ceiling.
Optane shoehorns its way into a memory hierarchy gap with SSDs below it and DRAM above it. Where will it gain most market share? Handy believes “that the 3D XPoint memory’s biggest impact will be to the server DRAM market.”
Customers get the highest application performance increase for their Optane cost that way.
Using Optane, servers can have a larger memory space as Optane DIMMs have four times the density of DRAM DIMMS. So servers can have more memory space. They can accommodate more or larger applications, which then run faster.
That’s because the apps avoid storage IO and therefore delays from IO requests passing through the storage software stack.
When Optane SSDs are used instead of NAND SSDs performance increase is limited to 5x to7x the SSD access speed because the storage software stack still has to be traversed by IOs. “For most applications the price premium is too high for the performance improvement,” Handy writes.
In other words Optane SSD prices have a lot further to go to cannibalise standard SSD sales.
DRAM DIMM and Optane DIMM price relationship
Handy notes that Optane DIMM prices are held to a constant relationship to DRAM DIMM prices.
Optane DIMMs are priced at slightly more than half the cost of half the DRAM capacity. High-density DRAM DIMMs cost more than lower density versions. and the same is true for Optane DIMMs. The logic is that Optane DIMMs are priced to replace DRAM.
When Optane is used instead of DRAM, applications (a) run a little more slowly because they have less DRAM, but (b) then run a heck of a lot faster because they have far less storage IO. Overall, the apps run much faster because (b) more than cancels out (a). That’s the theory.
Systems with Optane DIMMs must have some DRAM as well. Handy calculates that “At 6TB the memory in a DRAM plus Optane DIMM system costs 30 per cent less than a DRAM-only system.” With a 2X or more application speed up, depending on the Optane DIMM mode, that’s a worthwhile cost saving.
Why this is a problem for Micron
Micron makes and sells DRAM. Why should it displace a 256GB DRAM DIMM with a 256GB Optane DIMM when the DRAM DIMM brings in more revenue, and is likely profitable whereas Optane, being produced in relatively low volumes, is not?
We might expect Micron to focus on ways to sell Optane that don’t affect DRAM sales. For example, it may seek to displace Samsung or SK Hynix DRAM with its 3D XPoint DIMMs – branded Quantx.
Or, in the extreme case, abandon the Optane market altogether. Handy speculates this may be the case, with Micron having “desire to extricate itself from the 3D XPoint business while still satisfying alternate-sourcing agreements made by the company prior to XPoint’s 2015 introduction”.
Editor’s Note: This blog by storage industry veteran Hubbert Smith begins with a brief history of SLC-MLC-TLC-QLC and then gives a forecast of what to expect from QLC. It pertains mostly to server/cloud SSDs.
History
SLC – single level cell. It stores a single bit and a cell uses 2 charge levels. A full charge of 1.8V is a one and zero charge is a zero.
MLC – multi level cell stores two bits. A cell uses several charge levels to store two bits; 1.8V, 1.2V, .6V, and 0V.
TLC – three level cell stores three bits. A cell uses even more charge levels to store three bits; 1.8V, 1.5V, 1.2V, 0.9V, 0.6V, 0.3V, 0V.
QLC – quad level cell stores four bits. A cell uses even more charge levels… you get the idea. Voltage charge deltas are even tighter and more error prone.
Let’s look at the capacity and the value of the NAND and SSD:
As compared to SLC; MLC doubles capacity, adds 100 per cent more bits per cell.
As compared to MLC; TLC adds 50 per cent more bits per cell.
As compared to TLC; QLC adds only 25 per cent more bits per cell.
Diminishing returns
QLC is only 25 per cent better capacity than TLC, and with every generation the industry trades slower and slower performance with poorer write endurance. With just 25 per cent better capacity than TLC, QLC shows diminishing returns.
Additionally NAND is ugly and QLC NAND is a whole new level of ugly. Here’s why.
NAND cells are less than perfect. Firmware goes through all sorts of contortions to identify and correct media errors. As mentioned earlier, SSD data retention relies on the electrical charge of a cell and given enough time these electrical charges will evaporate. Recharging cells is one of the many maintenance tasks handled by firmware. When an SSD is plugged in, the firmware will refresh cell charges every 30 to 60 days.
What happens when an SSD is without a power source to refresh cell charges? No power means no cell recharge. Sooner or later the electrons will drift away and the cell electrical charge will evaporate; data loss occurs.
QLC promise and reality
The promise of QLC is very different than the reality of QLC. There are systems vendors attempting to drive to lower cost/GB. In a system this is likely workable.
There are memory vendors attempting to use QLC for cold storage; these folks are naively over-selling QLC.
Humble advice
Stick to proven TLC. Let someone else save a nickel and learn hard lessons.
Consider SLC for systems where the data sets are small but the over-writes are high. (Its endurance is far higher than that of MLC, TLC and QLC.)
Note: Consultant and patent holder Hubbert Smith (Linkedin) has held senior product and marketing roles with Toshiba Memory America, Samsung Semiconductor, NetApp, Western Digital and Intel. He is a published author and a past board member and workgroup chair of the Storage Networking Industry Association, and has had a significant role in changing the storage industry with data centre SSDs and Enterprise SATA disk drives.
Micron’s fourth quarter revenues were affected by the US trade dispute with China – and Huawei in particular. Quarterly revenues fell 42 per cent to $4.87bn, down from $8.44bn a year ago. Net income declined from $4.33bn to $561m.
Its outlook for the next quarter is $5.1bn at the mid-point, which contrasts with $7.9bn reported a year ago.
Micron CEO Sanjay Mehrotra said: “We are encouraged by signs of improving industry demand, but are mindful of continued near-term macroeconomic and trade uncertainties.”
You can be certain of that, with President Trump’s trade dispute with China nowhere near resolution. The Trump administration has barred sales of American software and componentry to firms on the so-called Entity list without an export licence.
Micron stated fourth quarter sales to Huawei – its biggest customer – were lower than anticipated. According to IHS Markit Huawei bought $1.7bn worth of DRAM and $1.1bn worth of NAND in 2018 – not just from Micron.
Micron has applied for licenses with the U.S. Department of Commerce to sell more products to Huawei. But the company said if U.S restrictions against Huawei continue it could see a worsening decline in sales to Huawei in coming quarters.
Q3 and Q4 fy19 revenue bars show slump is bottoming out.
Full fiscal 2019 revenues fell 23 per cent to $23.4bn and net income fell 55 per cent to $6.31bn.
DRAM sales accounted for 63 per cent of overall revenues and were down 48 per cent on the year. NAND sales, 31 per cent of overall revenues, declined 32 per cent.
Free cashflows were $263m for the quarter and $4.08bn for the full year. Micron has $9.3bn in cash, marketable investments, and restricted cash. Making DRAM and NAND eats capital, and Micron expects capital expenditures in fiscal 2020 to be between $7bn and $8bn.
Micron Technology President and CEO Sanjay Mehrotra said: “Micron delivered fourth quarter results ahead of expectations, capping a fiscal 2019 in which we executed well in a challenging environment, significantly improved our competitive position, and returned cash to shareholders through share repurchases.”
Q4 revenues for Micron’s four business units were:
Compute and Networking – $1.9bn revenues – down 56 per cent
Mobile – $1.4bn – down 26 per cent
Storage -$848m – down 32 per cent
Embedded – $705m – down 24 per cent
The DRAM and NAND downturns are ending, according to Micron. DRAM demand has bounced back as issues affected the first half of 2019 dissipated, while NAND elasticity – i.e. lower prices – is driving robust demand. NAND bit shipments will be higher than DRAM bit ships in the next quarter.
Micron chalked up record revenue and unit shipments in consumer SSDs and said it is positioned to gain share in the NVMe SSD market in fiscal 2020. It saw quarterly rises in demand for cloud and enterprise data centre products.
The company said it was on track to ship its first 3D XPoint products by the end of calendar 2019. That will signal the end of Intel’s monopoly in selling XPoint products.
Yellowbrick Data has introduced a public cloud version of its data warehouse, so customers no longer have to buy its custom scale-out analytics appliance.
The Cloud Data Warehouse is a SaaS offering that connects customers to a single Yellowbrick Cloud Data Warehouse via multiple public and private clouds. The service supports data warehouse volumes ranging from 10TB to several petabytes.
The Yellowbrick Cloud Data Warehouse has operated in enterprise production environments since early 2019. It now has the ability to replicate between the on-premises Yellowbrick Data Warehouse Appliance and the cloud, to support hybrid deployments on-premises with users in the cloud, and to move data to and from the cloud at will.
Cloud DR provides an up-to-date replica in the cloud of an on-premises or cloud data warehouse at a substantially lower cost than purchasing additional instances, according to Yellowbrick.
The Yellowbrick Cloud Data Warehouse is generally available with support for AWS, Azure and Google Cloud Platform. Yellowbrick Cloud Disaster Recovery will be generally available later this year.
Blocks & Files has been told by Yellowbrick that the Cloud Data Warehouse runs on Yellowbrick hardware in a co-location site with very fast direct connections to the public cloud. This means that Yellowbrick’s custom high-speed hardware is running the software and not the public cloud provider’s hardware. That may happen in the future.
About Yellowbrick
Yellowbrick Data was founded in 2014 by Chief Revenue Officer Jim Dawson, a previous chief sales officer at all-flash pioneer Fusion-io, and CEO Neil Carson. Carson is also an Fusion-io alumnus, having been its CTO in what now seems like the distant past.
CEO Neil Carson (left) and CRO Jim Dawson (right)
Its first data warehouse product, an integrated and scale-out analytics system, hit the market in September 2017. This all-flash system with custom hardware and software streamed data directly from the SSDs into the CPU’s cache.
The system used NVMe and bypassed main memory to get extra performance. This involved Yellowbrick writing its own operating system, schedulers, memory managers, device drivers and file systems. The claim was that users could get comparable performance to an in-memory database like SAP HANA, but with less server infrastructure.
Investors in both companies are no doubt attracted by the revenue potential for cloud data warehousing- projected by Allied Market Research projects to reach $34.7bn by 2025.
Intel showed off some performance figures for upcoming Alder Stream second generation Optane SSDs today at its Seoul Memory and Storage Day The info is a little sparse and presented in slide format. But it shows that Alder Stream should deliver at least 50 per cent better performance than the first generation P4800X Optane SSD.
Alder Stream is Intel’s codename for gen 2 Optane SSD using Barlow Pass 3D XPoint media with four layers. Current gen 1 Optane has two layers and is used in the DC P4800X SSD. The Alder Stream successor is due in 2020 and could have up to double the capacity of the P4800X. It uses a second generation controller which will make performance faster.
Three slides in Seoul showed relative performance indications:
ADS is Intel shorthand for Alder Stream
The first slide charts read latency as a function of total IOPS. Latency increases as IOPS rises. The NAND P4610 has the highest latency and latency increase on the chart, finishing at around 380,000 IOPS. The first generation Optane P4800X performs better, recording lower latency and giving up at 500,000 IOPS.
Alder Stream (ADS on the slide) starts at the same latency level but keeps on going consistently, rising slowly to reach at least 700,000 IOPS, the limit of Intel’s chart.
The second slide charts Alder Stream’s Aerospike benchmark performance:
The chart plots failure rate against device read latency and Alder Stream has the lowest failure rate compared to the 3D NAND P4610 and the Optane P4800X. The bar chart on the right shows maximum TPS (transactions per second) at less than five per cent failure rate.
The P4610 reached 300,000 TPS and the P4800X achieved 435,000 while Alder Stream soared to about 1.3m TPS.
The third slide examined the three drives on a Rocks DB test, looking at Gets/sec plotted against latency but with no actual numbers in the axes:
The hard-to-read slide above shows the P4610 curve on the left – the hard-to-see-number in the white circles are the number of threads. The P4800X is the middle line, riding steeply from 12 to 16 threads. Alder Stream is the lowest line and extends furthest to the right.
Intel said Alder Stream has three times the performance (Gets/sec) of the P4610 at the 16 thread count and four times lower latency. This misses the point somewhat. We are interested in how gen 2 Optane compares to gen 1 Optane, not how it compares to NAND.
How does Alder Stream compare to the P4800X? Based on the chart we think Alder Stream has 4.5 to five times lower latency at 16 threads than the P4800X.
Let’s award a nominal performance rating of 1 for the P4610. This makes Alder Stream a 3, and, by estimation the P4800X is 2. This gives Alder Stream a 50 per cent performance boost over the P4800X.
Overall, Alder Stream seems to have at least 50 per cent performance boost over the P4800X and slightly lower latency that is maintained over a wider workload range. Pair that with larger capacities and the gen 2 Optane SSDs will be big beasts indeed.
Hammerspace has introduced ransomware protection and other security services to its data-as-a-service platform.
The new features span hybrid and multi-cloud environments and are as follows:
Global undelete for files and snapshots that allows users to self-service data recovery
Automated data classification
Integration with customer-managed key management systems for multi-cloud security
Metadata harvesting integration with cloud analytics services to detect and tag files with content information
When files are deleted they are moved to the equivalent of a PC’s trashcan. They can be recovered – undeleted – using snapshot technology.
User-managed key management enables customers to put encrypted data in the public cloud, safe in the knowledge that it is secure against access, Hammerspace claims. Even if the cloud service provider suffers a breach, as AWS did in July this year with Capital One, the data is secure and the public cloud acts as a ransomware bunker, the company said.
Hammerspace provides a software control plane or gateway through a SaaS model, to access all of a company’s data, whether block, file or object, across all of its data centre sites and public cloud stores.
It is not a fan of NAS filers as an access path to files. CEO David Flynn told us in a briefing: “NAS as a delivery system for file is broken [because access and control are combined]. You have to be in the data path and that kills you.” Hammerspace separates control from the data path.
Hammerspace technology
According to Hammerspace, users and applications typically access data in specific silos in specific ways; data is stuck in place and format without dedicated tools to move and convert it. The control plane and data planes are co-located. Hammerspace Data Services technology separates the control and data planes. One generic storage and protocol-agnostic control plane can be used across all data stores, types and protocols, enabling data access planes to all of them.
Hammerspace Data Services diagram.
Hammerspace is developing a hybrid, multi-cloud environment covering on-premises bare metal, virtualized and containerised systems as well as AWS, Azure and GCP. Data lives in a universal namespace across all of these locations and can move between them.
Users and applications can access data anywhere within this namespace. Every data centre has a full copy of the metadata needed. The cloud-native Hammerspace Data Services software uses machine learning to optimise the system for performance and cost.
Kubernetes orchestrates container micro-services for users who declare what they need. Hammerspace orchestrates data services for users in the same declarative way, Flynn said.
Hammering away at the file access problem
Hammerspace emerged from the ashes of a defunct company called Primary Data and came out of stealth in October 2018. Flynn previously founded or co-founded FusionIO, bought by SanDisk for $1.1bn, and Primary Data. He told Blocks & Files ithat Hammerspace has done very well since its product was launched, earning “several millions of dollars in revenue”.
He said customers have had bad experiences with failed cross-data centre file virtualization technology and need convincing that Hammerspace works in the area where previous companies crashed out.
“Other suppliers’ messages about hybrid cloud data management are just fluff,” he said, citing recent acquisitions by Google’s Cloud business and Commvault. “These other suppliers weren’t successful in the market,” according to Flynn, implying Hedvig (Commvault) and Elastifile (Google) are hybrid cloud data management duds.
Intel’s second generation Optane storage-class memory is code-named Barlow Pass, and the company is developing 144-layer QLC (4bits/cell) NAND technology.
The chip giant is also opening a second Optane development line at its Rio Rancho, New Mexico facility, which will run gen 2 wafers through the first line.
At an Intel Memory and Storage Day briefing in Seoul today , Kristie Mann, head of product management for the Data Center Group, showed this slide;:
Mann’s slide shows there is a roadmap for generations 3 and 4 of Optane, with Sapphire Pass Xeons and then an un-named future Xeon after that.
The company will release Barlow Pass DIMMs in 2020, along with Cooper Lake (14nm) and Ice Lake (10nm) versions of the Xeon CPUs. Alder Stream gen 2 Optane SSDs are also coming.
A single-port gen 2 Optane enterprise drive is expected in 2020. A new controller will speeds things up even more.
Blocks & Files thinks Optane gen 2 will have 4 layers instead of the 2 in gen 1. We think DIMM gen 1 capacities of 128GB, 256GB and 512GB will double with gen 2 to 256GB, 512GB and 1TB.
Intel fellow Frank Hady confirmed this, and said gen 2 Optane has 4 layers, doubling the size of the die. In theory that means Intel could double the capacity of Optane drives and DIMMs.
More Barlow Pass details – capacity, endurance and performance – will come in a later article. Intel is in drip feed mode.
However the company revealed that Microsoft is updating its client OS to support Optane gen 2 features.
144-layer NAND
Intel will ship 96-layer QLC SSDs to customers next quarter. It will then move to 144-layer 3D NAND and will ship 144-layer SSDs in 2020. (This is likely to be a pair of 72-layer string stacks although we await confirmation)
VAST Data uses Optane and QLC flash. Expect VAST Data to use gen 2 Optane and the 96-layer and 144-layer QLC NAND drives. VAST’s Jeff Denworth, VP Product Management, said: “VAST is primed to go from today’s 64 layer QLC to whatever 3D NAND QLC offering Intel has in the future.”
Rob Crooke, GM for Intel’s Non-volatile Memory Solutions Group, said Intel is developing Optane persistent memory for PCs and notebooks that could remove the need for storage. Applications could be changed almost instantaneously by switching a pointer to Optane memory instead of loading the new app from storage.
Intel has launched the 665P – a faster version of its 660p quad-level cell NAND flash drive.
Blocks & Files saw the drive today at an Intel Memory and Storage day in Seoul, Korea.
The 660P is an M.2 format drive with a PCIe 3.0 x4 and NVMe v1.3. interface. It supports 0.1 drive writes per day (DWPDF) and has a five-year warranty. The drive uses SLC write cache and has 256MB of DRAM on board. It has random read/write IOPS numbers of 220,000/220,000 and 1,800MB/sec sequential read and write bandwidth. Here is a picture:
And here is the 665P, looking almost exactly the same;
Intel showed a CrystalDiskMark v7 benchmark test comparison of 1TB versions of the 660P and 665P.
660P
665P 1TB
The summarised numbers are;
660P sequential read/write – 1,229/1333 MB/sec
665P sequential read/write – 1,817/1,888 MB/sec
660P random read/write IOPS – 13,137/37,105
665P random read/write IOPS – 17.279/47,608
The 660P uses 64-layer NAND while the 665P uses 96-layers. Dave Lundell, head of strategic planning and marketing for Intel’s Non-Volatile Memory Solutions Group, said the media is better and the ASIC and firmware on the drive are also improved.
The 665P will hit the streets in the fourth quarter. No pricing information is available yet.
