Software-defined storage firm StorPool Storage has released version 19.3 of its scale-out block storage platform, adding management features to its GUI and a range of updates including broader compatibility with NVMe SSDs.
StorPool’s platform runs on Linux and pools HDD and SSD resources across a cluster of commodity server boxes to provide a shared storage system. The latest version offers greater support for cloud platforms based on the Linux KVM hypervisor, and for containerised workloads as well as those running in virtual machines.
In fact, StorPool claims that customers can now streamline their IT operations by operating a single StorPool storage system across all the cloud platforms they have, such as OpenNebula, CloudStack, OpenStack and Kubernetes for container orchestration.
The firm states that automation capabilities in the plug-ins for each cloud platform enable customers to manage their clouds from that cloud’s user console, so the underlying StorPool storage system will automatically perform any required storage layer actions. Each virtual machine or container deployed will get automatically dedicated volumes in StorPool, for example.
On the management side, administrators can now use the StorPool GUI to perform basic management tasks like creating and resizing volumes, creating snapshots, and others. Remote Bridge Status Monitoring has been added, showing the status of connections between a given StorPool cluster and other StorPool clusters, with analytics dashboards that show the bridge traffic between clusters over time.
This feature is to aid monitoring of multi-site deployments where many StorPool clusters replicate snapshots to one site, according to StorPool. It is also designed to help with large-scale deployments where tens of StorPool sub-clusters act as a single large-scale primary storage system presenting a unified global namespace.
In terms of integrations, StorPool has updated its support for OpenNebula, CloudStack, OpenStack and Kubernetes. The OpenNebula Addon now supports versions 5.12 and 6.0 of that platform, plus enhancements for security and to prevent data loss.
The CloudStack Plug-in adds support for CloudStack versions 4.11.3 to 4.15, and enables customers with CloudStack to use multiple StorPool Storage clusters as availability zones. With OpenStack, StorPool allows compute nodes to use the iSCSI protocol rather than the native StorPool Block Protocol, while admins can now take snapshots of Cinder volumes created from Glance images, and use them to deploy Nova instances. For Kubernetes, StorPool also now allows Kubernetes worker nodes to attach using the iSCSI protocol instead of the StorPool Block Protocol.
Hardware compatibility for NVMe SSDs has been expanded in this release by enabling storpool_nvmed to manage NVMe drives using the standard Linux vfio-pci driver, the firm said. StorPool has also confirmed compatibility with various NVMe SSDs, such asToshiba Cx5 NVMe, WD Ultrastar SN200, WD Gold NVMe WDS960G1D0D, HGST Ultrastar SN100, and Micron 9100 Pro.
We ran an email Q&A with Young-Soo Sohn, Corporate VP/Head of DRAM Memory Planning & Enabling Group, Samsung Electronics, concerning DRAM, High-Bandwidth Memory (HBM) and Storage-class Memory (SCM) developments and received fulsome answers concerning DRAM and HBM. Less so with SCM which is more of an emerging technology than the other two technologies.
DRAM
Blocks & Files: What memory nodes is Samsung supporting? 1Z, 1alpha, 1 beta, 1 gamma? I ask this because Micron is making a thing about future memory node support.
Young-Soo Sohn: We recently announced the industry’s most advanced 14nm DRAM based on extreme ultraviolet (EUV) technology*. By increasing the number of EUV layers to five, we were able to create the smallest possible DRAM node today, which will enable unprecedented speeds. To capitalize on this, we plan to mass produce 14nm-based DDR5 in the second half of this year.
Further, we are developing next-generation process nodes to accommodate the industry’s most demanding applications, all of which require improvements in density, performance and power consumption.
*Samsung was the first in the industry to adopt EUV in DRAM production. EUV technology reduces repetitive steps in multi-patterning and improves patterning accuracy, resulting in increased performance and shortened development time.
What are Samsung’s plans with regard to DDR5?
We have already provided our customers with samples and fully intend to fulfil their needs as DDR5 mass production launches in the second half of this year.
Samsung is working closely with key industry stakeholders to deploy DDR5 product in various high-performance applications. We are currently sampling different variations of our DDR5 family to customers for verification and soon for certification with their leading-edge products to accelerate AI/ML, exascale computing, analytics, networking, and other data-intensive workloads.
When does Samsung think DDR6 technology will emerge? What advantages will it bring to systems?
While we are not in a position to detail our DDR6 plans at this point, be assured that we are committed to delivering high-performing DDR6 modules in a timely fashion once the market is ready.
Broadly speaking, the computer industry is fast-moving and DRAM will continue to play a central role in moving the industry forward. As computing power continues to increase, DRAM performance will also need to improve to keep pace. From this standpoint, the transition to DDR6 will be inevitable, and we are partnering with global manufacturers and investing in cutting-edge technologies to ensure high-performing, energy-efficient memory solutions that will accelerate AI/ML and other compute-intensive applications. Without question, DDR6 will be the next key player and we intend to remain at the forefront of the industry when this transition occurs.
HBM
What advantage in capacity terms would HBM supply over DRAM?
This varies by generation. The current HBM2E standard (an extension of second-generation HBM2) supports stacking of up to eight stacks, which would allow an HBM2E SiP to provide 128GB of capacity. In practice, however, currently available HBM2E capacities top out at 16GB. The HBM3 standard (now being developed) is expected to extend stacking capabilities — in conjunction with increasing device densities, this will produce significantly higher maximum capacities.
Samsung table.
What advantage in bandwidth terms would HBM supply over DRAM?
This is where HBM’s advantages are most apparent in terms of bus width and processing speed. Since their first generation, HBM standards have supported a 1024-bit bus width, compared to just 32 bits for GDDR, and the standard HBM2E signaling rate is 3.6Gb per second (Gbit/sec) per pin, with up to 460GB per second (GB/sec) of bandwidth per stack. At this point in the development of HBM3, we expect to achieve the ever-fast processing speed of up to 6.4Gbit/sec — well above many expectations.
Unlike the case with capacity, however, in practice bandwidth has gone beyond the standard — Samsung’s HBM2E Flashbolt devices have a processing speed of 3.6Gbit/sec per pin, and bandwidth up to 460GB/sec. This compares to standard-specified data rates of 16Gbit/sec for the GDDR6 generation of DRAM. And for processors with a 4096-bit memory interface, such as GPUs and FPGAs, a combination of eight Flashbolt stacks can offer 128GB of memory with 3.68TB/sec peak bandwidth — far beyond traditional DRAM.
How would HBM be constructed in terms of, for example, number of stacks (layers) of DRAM and an interposer and a CPU?
As noted above, the current HBM2E standard supports up to eight dies each, while the as-yet-unreleased HBM3 standard is expected to increase the maximum number of dies.
Use of through-silicon vias and microbumps (rather than wire bonding) for interconnection of the stacked memory dies allows for a very small footprint and fast data transfer rates, while also enabling improved heat distribution to minimize thermal issues.
Moreover, use of a silicon interposer for interconnection of HBM stacks and processors means that the memory and processor can be in very close proximity for reduced access time, taking advantage of the high efficiency of silicon-based interconnections. The end result is both improved performance and substantially reduced energy consumption compared to the board-level interconnections typically used for traditional DRAM. And going forward, the interposer offers opportunities for inclusion of additional active circuitry that could advance the HBM model from its current 2.5D status to true 3D integration [including the very promising processor-in-memory [HBM-PIM] avenue].
Which supplier affixes the CPU to the HBM?
Assembly and test of the HBM and processor requires advanced techniques, but they are similar to those that are increasingly being used in chiplet-oriented designs where multiple dies are connected within an IC package. Each system OEM will make its own manufacturing supply chain decisions, but it seems now that members of the existing outsourced semiconductor assembly and test (OSAT) community (ASE, Amkor, etc.) will have the necessary capabilities as will some integrated device manufacturers (eg. Intel) and foundries (eg. TSMC).
Which CPUs would be supported? Which Xeon generation? Which AMD generation? Arm? Risc-V?
We are working with a variety of CPU developers. For ARM, Fujitsu is supplying and Xeon/RISC-V is currently developing.
Would/could HBM be connected to servers across a CXL bus? Is this a good or a bad idea?
We are deeply involved with the Compute Express Link (CXL) interconnect standard and continue to closely monitor its development. As with most technologies, the role of a specific memory technology will be determined in many ways by application requirements.
Will systems (servers, workstations, laptops) use both DRAM and HBM together? Why would they do this?
New products and projects have always presented system designers with unique sets of considerations on performance, power consumption, cost, form factor, time to market, and more. Today, with computing technology clearly in an era of unprecedented advancement, the calculations and tradeoffs are more complex than ever — with AI, ML, HPC and other emerging applications driving change at every level, from supercomputing centers to edge devices.
In response to these dynamics, and with traditional computing models and architectures under increasing pressure, engineering expectations for memories are rapidly evolving and diverging. As a result, an increasing array of memory types (GDDR6, HBM, DDR5, HBM+PIM, etc.) is now being used, especially in the HPC world, and this increasing heterogeneity is likely to extend to almost every market sector over time, as application demands grow more specialized. Again, these options are enablers for creative, innovative engineering and we expect designers to take advantage of all available options to meet their specific objectives.
Storage-class Memory (SCM)
What is happening with Z-NAND?
We are carefully reviewing the development direction of Z-NAND by collecting customer requirements.
If HBM is taken up by customers will there still be a need for SCM in the memory-storage hierarchy?
With the explosive growth in data generation, the need for delivering data-centric solutions for various workloads has never been greater. We are preparing solutions in close cooperation with a variety of partners to meet these radical demands for memory innovation.
What SCM technologies have attracted Samsung’s attention and why?
There are various workloads in the industry with different requirements, which should be addressed accordingly.
How does Samsung view Intel’s 3D XPoint technology?
As mentioned above, there are various needs in the industry, and a variety of solutions to address those needs.
Comment
Samsung is energetically active with DDR5 and DDR6 memory. It is also strongly developing its HBM and HBM-PIM stacked memory technology products. We note it’s thinking about HBM with X86, Arm and RISC-V processors. HBM-Arm and HBM-RISC-V would provide interesting competition for Xeon-HBM systems, particularly in edge environments with low-latency, data-intensive processing needs.
The Storage-class Memory situation is less clear-cut. It appears that Samsung sees a general need for SCM technology, but it’s possible application-specific technology may be needed rather than a single SCM technology.
Ransomware rules the waves in this digest. Prepare to be numerically amazed by the number of ransomware attacks and then read good news about Active Directory recovery from ransomware. We also cover Delphix’s sustainability initiative and an InfiniBand spec update.
Ransomware prevalence
According to the data presented by the Atlas VPN team, ransomware attacks are up 151 per cent in the first half of 2021, compared to the same time in 2020. The United States is the most ransomware-threatened country by a significant margin.
The report shows that:
In the first half of 2020, there were a total of 121.4 million ransomware attacks across the world.
The first half of 2021 has already reached 304.7 million ransomware attack attempts, making it the worst year ever recorded.
The United States is a prime target for most hackers, as the country has experienced 227.3 million ransomware attack attempts.
The second most targeted country is the UK, as it encountered 14.6 million ransomware attack attempts.
We find these numbers simply amazing. There is a concerted and sustained ransomware Blitzkrieg (or should that be Mолниеносная война) going on.
Quest updates Active Directory DR
Quest Software has updated its Disaster Recovery product which has a very long name: Recovery Manager for Active Directory (RMAD) Disaster Recovery Edition (DRE). We are highly amused by Quest’s use of two separate acronyms — RMAD and DRE — and agree that combining them into RMADDRE looks, well, crazy. So RMAD DRE it shall be.
RMAD DRE v10.2 adds functionality to protect Active Directory (AD) backups from malware, and the ability to restore AD to a clean Microsoft Azure virtual machine in the cloud.
Michael Tweddle, President and General Manager, Quest Microsoft Platform Management, said: “Ransomware is the number one security threat today, with 69 per cent of businesses suffering an attack in 2020 alone, and an average downtime of 21 days. Following an attack, business operations cannot be restored until Active Directory is functional again.”
Many ransomware strains like SaveTheQueen and DopplePaymer compromise AD and leverage it to spread throughout the target organisation’s systems.
The Secure Storage functionality provides a hardened server that is isolated from the network according to IPSec rules with regular checks to confirm backup integrity.
Delphix and sustainability
Data ops supplier Delphix says it will increase its corporate commitment to environmental sustainability.
Jedidiah Yueh, founder and CEO at Delphix, said: “Climate change is a preventable crisis. Our platform helps customers reduce their data footprint and carbon impact by 10x for enterprise applications, so environmental sustainability is core to our value as a company.
“We are building a sustainability strategy that includes creating a Sustainability Board, appointing a Director of Sustainability, measuring our company and platform impact, and collaborating with our customers to achieve common environmental goals.”
Delphix says this adds to the company’s foundation of social giving and its diversity, equity, and inclusion efforts. It has a Diversity, Equity, and Inclusion (DE&I) Council with several Employee Resource Groups (ERGs) catering to specific interest groups. Since 2015, these ERGs have led company-wide programs to celebrate diversity, including International Women’s Day, Juneteenth, Pride, Asia-Pacific heritage month, etc.
The company recently announced an increase in its overall employee diversity rates from 29.5 per cent in Q4FY21 to 31.7 per cent in Q1FY22 — a growth of over two per cent over the previous quarter. It also saw an increase in its female workforce from 22.8 per cent in Q4FY21 to 24.7 per cent in Q1FY22 — nearly two per cent up on the last quarter.
Supermicro’s storage boxes
Server and storage server supplier Supermicro has announced top-loading (60-bay and 90-bay) and Simply Double storage systems.
Both 60-bay and 90-bay top-loading architectures are optimised for enterprise and cloud-scale storage environments, and available in single-node, dual-node, and dual-node high availability (HA) configurations.
The dual-node HA and single-node configurations control access to all drives in the systems. The dual-node configuration evenly split the drive control access between each node. With a modular, tool-less design, all critical onboard systems — hot-swap server nodes, expanders, fan modules, power supplies, and drives — are fully optimised for easy serviceability by one technician.
Supermicro’s top-loading and Simply Double enclosures.
These 4U systems feature 60 or 90 hot-swap 2.5-inch or 3.5-inch SAS3/SATA3 bays plus two onboard PCIe M.2 slots and two internal slim SATA SSD slots. The single-node system also supports two rear hot-swap 2.5-inch bays for OS mirroring and four optional NVMe U.2 bays for fast caching. The system supports 1.6 petabytes of cost-optimised storage at the maximum configuration, plus up to 60TB of SSD flash via the rear-accessed NVMe compartment.
The single- and dual-node systems use 3rd-Gen Intel Xeon Scalable processors in a dual-socket configuration with 16 DIMM slots per server node.
Simply Double storage systems promise to enhance the serviceability and performance of that design, while maintaining the same dense storage footprint. Up to four rear hot-swap U.2 NVMe bays are supported, allowing users to add flash without sacrificing any of the 24 SAS/SATA 3.5-inch storage bays. The system can be configured with the options of PCIe 4.0-based RAID or IT mode SAS controller.
The systems support PCIe Gen-4 or SAS. Supermicro says they are good for cloud-scale and HPC workloads.
InfiniBand Trade Association updates spec
The InfiniBand Trade Association (IBTA) announced IBTA Specification Volume 1 Release 1.5, with support for Memory Placement Extensions (MPE) to significantly reduce latency in Persistent Memory (PMEM) applications.
The specification includes key features that address industry demand for higher bandwidth and lower latency connectivity:
Memory Placement Extensions — enable hardware OEMs and software developers to implement the MPE operations RDMA Flush and RDMA Atomic Write and reduce overall latency by 10x;
Support for NDR — allows 400 Gbit/sec speeds on a four-lane port configuration.
Quality of Service enhancement via an updated VL Arbitration mechanism — allows a user to specify minimum bandwidth on physical ports and control virtualized port arbitration by setting rate limiting and bandwidth sharing
Shorts
Aerospike ran a petabyte-scale benchmark of its Real-time Data Platform delivering four million to five million transactions per second (TPS) with sub-millisecond latencies for read-only and mixed workloads on a 20-node Amazon Web Services (AWS) cluster using Xeon SP CPUs. It claims the Aerospike Database performs with hundreds of nodes less than other databases, saving up to $10 million per application in infrastructure costs.
AWS announced general availability of Amazon MemoryDB for Redis — a fully managed, Redis-compatible, in-memory database. It is claimed to enable customers to achieve ultra-fast performance with high availability and durability for their business-critical applications that require sub-millisecond response times.
Cisco results for its Q4 FY2022 quarter showed its Infrastructure Products group revenues were $7.55 billion, up 13 per cent year-on-year. All businesses in the grew double digits except for the data centre segment, which declined due to weak server revenue. Management attributed this to server market contraction but William Blair analyst Jason Ader suspects Cisco market share loss might be a factor as well.
Cloud-native data linking Confluent has announced its Q321 release of software, its first quarterly release. It adds data mirroring to cluster linking for near-time DR, ksqlDB pull queries to simplify stream processing architectures, fully managed connectors for Salesforce and Azure, and Infinite Storage for Google Cloud, after initially launching for AWS.
Fermín Serna.
Incredibly well-funded Databricks has hired a Chief Security Officer, Fermín Serna. He comes from Citrix with a near-eight-year stint at Google working on security before that. He will will lead Databricks’s high-impact network, platform and user security programs along with governance and compliance efforts.
SaaS data protector Druva has hired Ash Parikh to be its CMO. He comes from Informatica where he was an SVP and Global Head of Marketing. Parikh replaces Thomas Been, the previous CMO at Druva, who has gone off to join DataStax as its CMO. Parikh, now a ‘Druviuan’, has written a blog about his first few days at the company.
An independent evaluation and analysis by Customer Relationship Management Institute LLC (CRMI) has certified Druva’s record-high NPS score of 89 and has awarded the company the NorthFace ScoreBoard Award for the third consecutive year.
SaaS data protector HYCU has added Devoteam to its Cloud Services Provider Program. Devoteam operates in EMEA and is a Premier Google Cloud reseller and services partner. Joint customers adopting Google Cloud now have a cost-effective and easy way to store on-premises backups and an efficient DR platform.
Mainframe VTLer Model9 has bought a z15 mainframe. Gil Peleg, Model9 CEO, said: “The Model9 z15 will be a powerful asset for our product development and quality assurance programs.” Model9 says it has been growing massively through the pandemic period, tripling its size both in number of customers and headcount, as organisations seek to maximise the value of their mainframe investments while also leveraging the capabilities of the cloud.
The Nutanix Certified Professional — Multicloud Infrastructure certification is eighth in the top 15 list of top-paying certifications by Global Knowledge. That means it is amongst the most in-demand skills and technology areas for organisations today and, crucially for employees, the average salaries associated with them.
Nutanix and Cyxtera have partnered to launch the first Nutanix Federal Innovation Lab, powered by Cyxtera’s digital exchange and Enterprise Bare Metal. The lab, located in a Cyxtera data center in Northern Virginia, provides US Federal customers as well as industry partners with an environment to build proofs of concept and test mission-critical applications using Nutanix software.
PNY has announced Performance DDR5 4800MHz Desktop Memory for desktop PC and gaming systems. Available in the third quarter, it will start at 16GB per module and feature the JEDEC standard frequency of 4800MHz. DDR4 operates at 3200MHz. Intel’s Alder Lake CPUs launching later this year will be the first to support DDR5 memory. DDR5 will likely overtake DDR4 in popularity by 2023.
Spectro Cloud announced the release of an open-source CNCF Cluster API contribution supporting Canonical’s MaaS interface. It enables organisations to deploy, run and manage Kubernetes clusters directly on top of bare metal servers, increasing performance and minimising cost and operational effort.
Storage conglomerate StorCentric has signed up TechnoBind as a value-added distributor (VAD) in India. TechnoBind will supply StorCentric’s Drobo, Nexsan, Retrospect, Vexata and Violin Systems products in its territory.
Storage stack rewriter and accelerator StorONE has signed a strategic distribution agreement with Daiwabo Information System Co., one of Japan’s leading technology distribution firms. DIS will be the first firm in Japan to offer StorONE’s pioneering platform for deploying any combination of workloads (file, block, object), supporting all storage protocols, and utilising any mix of media and applications.
Intel has announced two more IPUs and a SmartNIC at an Architecture Day event, as it disses the DPU idea and doubles down on the IPU term.
These Intel IPUs (Infrastructure Processing Units) process IT infrastructure-related events for cloud service providers, as opposed to their tenant CPUs which process client applications and directly generate revenue. Intel’s first publicly revealed IPU is code-named Big Springs Canyon. It is FPGA-based and features a Xeon-D CPU, and Ethernet support.
Intel’s Patricia Kummrow, VP Network and Edge Group inside the GM Ethernet Products Group, has blogged about these three new products, writing: “CPU cycles spent on … infrastructure overhead do not generate revenue for the Cloud Service Providers (CSP).” Offloading infrastructure overhead from the (predominantly X86) tenant CPUs to specialised Infrastructure Processing Units (IPUs) means the work can get done more efficiently, and revenue from extra client work done on the tenant CPUs exceeds the IPU cost.
IPU canyoneering
The Big Springs Canyon successor is called Oak Springs Canyon, and is based on Intel’s Agilex FPGA (download a Product Brief [PDF]) and a Xeon-D SoC. Intel claims this is the best FPGA in existence in terms of performance, power consumption, and workload efficiency. It supports PCIe Gen-4, 2 x 100Gbit Ethernet networking, and Intel’s Open FPGA Stack hardware and software can be used with it to develop applications.
Intel’s second new IPU is called Mount Evans and employs an ASIC instead of an FPGA. In general, FPGAs are multi-purpose, programmable and flexible, whereas ASICs feature silicon-based hardware designed for a specific application. They should be faster than an FPGA used for the same application.
The Mount Evans ASIC is designed for packet processing, which is also an aspect of the Fungible and Pensando DPU chips. Intel claims Mount Evans, co-designed with a CSP, is “a best-in-class packet-processing engine” but doesn’t say what that class is. Mount Evans has up to 16 Arm Neoverse N1 cores. It can support up to four host Xeons with 200Gbit/sec full duplex. All this means it supports vSwitch offload, firewalls, and virtual routing and has headroom for future applications.
Intel says it “emulates NVMe devices at very high IOPS rates by leveraging and extending the Intel Optane NVMe controller”. That sounds, well, impressive, but it’s hard to know what it means in practice. A slide note says it is scaled up from the Optane implementation — which suggests it goes faster.
Both Oak Springs Canyon and Mount Evans run the same Intel IPU operating system.
SmartNIC
The N6000 is called an Acceleration Development Platform and is also based on the Agilex FPGA plus Intel’s E810 Ethernet controller. The N6000, code-named Arrow Creek, is based on Intel’s N3000 FPGA Programmable Acceleration Card (Intel FPGA PAC) N3000. This is deployed in Communications Service Providers’ (CoSPs) data centres.
The N6000, which supports PCIe Gen-4, is designed to help telco providers accelerate Juniper Contrail, OVS, SRv6 and similar workloads. It and the DPUs work with, naturally, Xeon CPUs.
Comment
Intel has not supplied any performance numbers enabling comparisons to be made between its three IPUs, nor between them and DPUs from Fungible and Pensando. N6000 performance numbers are not publicly visible either.
Intel says the DPU term is nonsensical, as a Xeon CPU processes data. In its view, it’s far better to call these accelerators IPUs because that’s what they accelerate: infrastructure-oriented processing.
Bloomberg reports unrevealed sources saying Databricks is raising a $1.5 billion-plus round, valuing the company at $38 billion.
This follows on from a $1 billion G-round fund raiser in February which valued Databricks at $28 billion. Are we seeing a price bubble?
Databricks provides analytical services on a data lake — raw data stored before being extracted, transformed and loaded into data warehouses such as Snowflake. Its open-source Delta Lake software is built using Apache Spark and data is processed using SQL Analytics software.
Taking this in-process funding round into account, Databricks will have raised $3.4 billion in total.
We have called this latest round an H-round in our table, as it uses a new and higher valuation for Databricks.
We can’t see how Databricks needs the money. It’s inconceivable it could have spent a billion dollars to grow its business since February.
We all know about price bubbles — when people, including investors, pay way over the realistic price for a stock. Is it possible that savvy VCs, encouraged by Snowflake’s massive and successful IPO, are pouring far too much money into data analytics startups and won’t get their money back? Are the data analytic startup valuations going so high as to be fantasies?
CTERA’s CTO, Aaron Brand, declares AWS S3’s pricing scheme “is anything but simple” and “frustratingly expensive” to transition your objects between storage classes. In fact, “this complexity is driving their customers mad”.
Update. CTERA has updated its S3 storage costs blog to reflect S3 price model changes announced on AWS Storage Day 2021. That makes the comparison data here invalid. Check out the CTERA blog for the new comparisons. 15 Sep 2021.
These are the concluding remarks in a blog in which he discusses S3’s three main storage classes, recurring fees and per-operation fees.
* Assuming the object is uploaded to “S3-Standard” and then transitioned using storage lifecycle.
The three classes are: Standard, Infrequent Access, and Intelligent Tiering. Brand says: “All these storage classes offer identical performance, so, the choice between them does not depend on the level of performance.”
Aaron Brand.
In Brand’s view, “S3 standard is always more cost effective in storing objects smaller than 128KB. If your application stores many small objects, keep those in S3 standard.“
Although Infrequent Access is cheaper than Standard, the Put, Get and Retrieve operations cost more. Brand has an ROI model to calculate whether it is worthwhile keeping a group of objects in Standard or Infrequent Access.
This takes into account:
The upfront cost (the cost of the PUT operation and lifecycle transition);
The fixed monthly ongoing costs (the per-GB cost);
The variable costs (the cost of data retrievals).
He worked out the costs and ROI in months for sets of objects with average object sizes of 128KB, 250KB, 1MB and 4MB:
In his view, “As a rule of thumb, for objects of 1MB or larger, S3-IA is recommended if the number of monthly retrievals per object is 0.7 or less. For smaller objects, be very careful, as anything more than one yearly retrieval may result in a negative return on investment.”
Brand says “S3 Intelligent Tiering charges you a fixed and predictable management fee based on the total number of objects. This management fee of $2.5 per million objects, per month, is insignificant when dealing with large objects but may be cost prohibitive for smaller objects.”
Brand concludes that the S3 pricing scheme is expensive, and working out costs of transitioning objects between storage classes is frustrating.
The main takeaway is that the important thing for keeping S3 costs as low as possible is not using it to store small objects. “By putting high price tags on metadata operations, AWS is severely penalising workloads consisting of numerous operations on small objects.”
His recommendation is to keep your S3 stored objects 1MB in size or larger.
Kumar Sreekanti, HPE’s CTO, head of software, and SVP, is retiring as HPE implements a GreenLake cloud services development scheme and team. There will be a new CTO appointed by the end of September, and Sreekanti will stay in place until then.
This was announced in a blog by HPE CEO Antonio Neri. He says HPE is transforming into an edge-to-cloud company, and “I am making this transformation my number one priority, and I will devote a significant amount of my time and energy to propel our strategy forward”.
Neri’s organisational transformation
Antonio Neri.
Neri wants HPE to transform more quickly and is setting up:
GreenLake Platform Development team to be fully accountable for delivering our new unified GreenLake Cloud Platform where all of our edge-to-cloud services will be delivered. This team will report to the CTO’s office to ensure it is delivering a holistic HPE as-a-service experience across all as-a-service businesses.
GreenLake Cloud Services Solutions group to be led by Vishal Lall reporting directly to Neri. The group includes the Ezmeral, GreenLake Lighthouse and workload offerings and will delivering services that integrate HPE’s unified edge-to-cloud platform.
Keith White will lead the GreenLake Cloud Services Commercial Business group, which is focused on driving the HPE as-a-service go to market through HPE’s Sales, Partner and Customer Success functions.
There are two new company-wide teams — the Services Reliability Engineering (SRE) team and the Cloud Services Operations team — which will be responsible for running the cloud platforms, services and security functions, and operations across HPE’s as-a-service offerings. These teams will report to COO John Schultz, who manages the HPE’s Transformation Office.
Kumar Sreekanti
Neri writes: “These updates more closely align key resources and teams to a Cloud Services business and delivery model, all designed to accelerate the delivery of a seamless, unified, and agile customer experience, from edge-to-cloud.” He pitches HPE’s GreenLake as the “cloud that comes to you.”
Sreekanti
Sreekanti joined HPE when it acquired BlueData in 2019 with its EPIC container environment deployment platform. Sreekanti was BlueData’s co-founder and CEO. EPIC was a control plane for container management and became part of HPE’s Ezmeral portfolio in June 2020.
Has Sreekanti become redundant in Neri’s drive to transform HPE? Clearly he does not want to be the new CTO — or Neri thinks he isn’t best suited to be. Perhaps Sreekanti has a new opportunity in mind?
HPE has also hired Sheila Rohra away from NetApp to be the SVP and General Manager of its new Data Centre Infrastructure business, reporting to Storage SVP and GM Tom Black.
SK Hynix has doubled the capacity of its Gold P31 NVMe gumstick SSD by using terabit dies instead of the original 512-gigabit ones.
The dies use 128-layer TLC (3bits/cell) flash and the drive has an SLC cache, like the original P31 products.
Like the original P31 drives, this new one comes in the single-sided M.2 2280 format and is intended for notebooks used for things like 3D design work and other content creation activities, and gaming systems. There is no improvement in performance over the original P31 drives, but the endurance increases from the 1TB model’s 750TB written (0.4 drive writes per day) to 1200TB written, or 0.3 drive writes per day for the five-year warranty period.
Performance summary:
Random read IOPS — 570,000 using SLC cache, 500,000 if cache full;
Random write IOPS — 600,000 with cache, 370,000 without;
Sequential read — up to 3500MB/sec with cache, 1700MB/sec without;
Sequential write — to 3200MB/sec with cache, 950MB/sec without;
Interface — PCIe Gen-3 x 4 lane;
Active power — 6.3W;
Idle power — <50mW;
L 1.2 idle — <5mW;
Controller — SK hynix Cepheus.
A 2TB Platinum drive concept was floated by SK hynix when the 512GB and 1TB Gold P31s were launched a year ago. That has come to naught and the 2TB drive is just an upgrade refresh for the Gold P31 brand.
If SK hynix cared to do so it could make a 4TB Gold P31 drive by using both sides of the M.2 format card. We might expect it to bring out PCIe Gen-4 versions of the drive in the future, with approximately doubled performance. A 4TB PCIe Gen-4 gumstick drive would be a neat way of giving a notebook computer a great big storage IO boost.
The company has upped its 3D NAND layer count from 128 to 176 and this could be used to increase the drive’s capacity as well, or to keep it the same and reduce its cost.
To make it a tiny bit special the 2TB Gold P31’s backing printed circuit board is coloured black instead of the original product’s green. The price is $279.99 with a 20 per cent off deal at Amazon on August 23.
Tape media sales dropped last year — but the hyperscalers are buying more tape libraries this year, which should lead to a tape media sales rise as night follows day.
A report by the The LTO Program Technology Provider Companies (TPCs) — Hewlett Packard Enterprise, IBM Corporation and Quantum — indicates that 105,198 petabytes of total tape capacity (compressed) shipped in 2020. The report says this was “slightly short of 2019 (a record year)” and that “for context, real GDP fell to a -5 per cent growth rate during the same time period”.
Eric Bassier, Senior Director, Quantum, issued a statement: “Coming off record capacity shipped in 2019, we were optimistic for 2020 — but global shutdowns and other factors outside of our collective control led to a reduced performance. We’re optimistic that there will be a return to the prior capacity growth trend in 2021.”
He opined that “Despite the unexpected headwinds for many segments of the technology industry produced by the pandemic, overall LTO tape capacity shipped in 2020 was strong in context.”
The report does not actually reveal though, that tape media sales fell 7.8 per cent from the 2019 total of 114,079PB — more than the five per cent GDP drop — as a look at the 114,079PB 2019 number reveals.
The TPCs issued a chart showing the 2019 number, but have not issued one showing the 2020 number. We took the TPC-issued chart for shipments to 2019 and added a column for 2020 plus a few annotations:
Blocks & Files additions to 2020 LTO Tape Provider Company chart
As you can see, the 2021 number was below both the 2019 and 2017 total tape media sales numbers — possibly a little bit more than a “slight drop”.
Quantum is reporting increased tape library system sales to hyperscalers, so tape media sales should start picking up this year and into next year. Even so, the 2021 total media ship number may not equal the 2019 record.
A snapshot of the IT environment provided by a Veeam survey shows that hybrid IT has become reality, cloud disaster recovery is difficult and a minority interest, and Office 365 data needs backing up — as does cloud-native app data.
Data protector Veeam has issued a 2021 Cloud Protection Trends Report based on data from a survey run by an independent research firm. It looked at data protection in four areas of IT: hybrid cloud, disaster recovery to the cloud, SaaS app protection and container protection.
There are three highlighted results”
Almost half of the 1550 respondents run production apps in the public cloud;
80 per cent use the public cloud in their disaster recovery (DR) arrangements;
Twice as many organisations use a third party backup for Office 365 compared to a year ago.
The survey showed that hybrid IT — the mix of on-premises and public cloud IT environments — is a reality. Backup is changing from reliance on physical and virtual servers in the data centre to increased reliance on hosted virtual machines in the public cloud.
Veeam survey report chart.
More than half (55%) of the respondents used the public cloud for normal production workloads, 47 per cent for high-priority production workloads, 36 per cent for development and 21 per cent for disaster recovery.
Application movement is not one-way to the cloud. Some 23 per cent of respondents have decided to bring on-premises applications moved to the cloud back home. That movement reinforces a requirement that data protection facilities should cover both the on-premises and the public cloud worlds (AWS, Azure, GCP, etc.), and the environments within them — virtual machines and containers.
The public cloud is popularly used for DR, with less than a fifth of respondents (10%) not using it. For example, 40 per cent store backed up data in the cloud, for on-premises restoration, 39 per cent use the cloud as a secondary site for DR, and 27 per cent have a purpose-built DR-as-a-Service arrangement involving the public cloud.
Respondents using the public cloud in their DR arrangements pointed out there were difficult areas that could be smoother — network configuration and connectivity, securing remote sites, firing up servers and verifying remote server functionality. Expense was another issue.
On the more positive side, there was widespread acceptance that data used by SaaS applications, like Office 365, needs to be backed up by the SaaS user. This is to protect against accidental deletion, cyber and other malicious attacks, and to meet compliance needs.
There was also acknowledgement that containerised apps and data need protecting. In both this and the SaaS application case, the decision to protect and to manage the protection was split between backup admin staff and SaaS/PaaS admin staff.
Comment
In the world of hybrid IT, apps and data are running on physical, virtual and containerised servers in data centres, as well as on virtual and containerised platforms in the public clouds, and also accessed as services from the cloud. This inherently makes the provision of a single backup environment covering all three environments a formidable task. Even moreso if restoration is desired outside the source environment.
Ideally there should be a single or unified data protection control plane. This would enable different employees, backup admins, SaaS/PaaS admins, DevOps, compliance and line-of-business people to co-operate, co-ordinate and integrate their data protection needs.
You can download a copy of the PowerPoint deck-style 14-page report from Veeam’s web site.
A clustered 16-controller Inspur storage array system using Optane and NVMe SSDs has set an SPC-1 performance world record, overtaking Huawei.
The Storage Performance Council’s SPC-1 benchmark tests the performance attributes of a storage array responding to business-critical workload IO requests. It measures SPC-1 input/output operations per second (IOPS), price per IOPS, response time, capacity and price per gigabyte.
Inspur’s HF18000 G5-1 system scored 23,001,502 SPC-1 IOPS, beating the previous record set by Huawei of 21,002,561 IOPS. The Huawei Dorado 18000 v6 array used 576 x 1.92TB NVMe SSDs and Inspur’s array had 570 SDDS as well. But that number included 480 x 1.92TB NVMe SSDs plus 96 x 375GB Optane SSDs as well. The P4800X Optane drives tipped the balance.
A chart shows SPC-1 ranking by IOPS;
Fujitsu was first to push performance past 7.5 million IOPS, but Huawei blew that away with its all-NVMe SSD system. And now Inspur has used Optane acceleration to overtake Huawei.
There are no Western suppliers in the top 10 SPC-1 rankings:
This benchmark is largely a game played by Chinese suppliers now.
Plotting the top results in a 2D space defined by IOPS and $/IOPS we can’t see a general trend of lowered cost per IOPS as performance increases:
The chart shows that Inspur systems do tend to be lower cost than Huawei ones.
Will Western suppliers of external storage arrays — Dell, HPE, Infinidat, NetApp, Pure and VAST Data for example — bother with SPC-1 benchmarks anymore? There is little to no indication that their marketing of product performance needs an SPC-1 boost. And the SPC-1 test does not reflect NVMe-over-Fabrics technology, nor NVIDIA’s GPU Direct for that matter. In this sense it is yesterday’s benchmark and needs replacing.
Having said that, this is a great result for Optane — Intel will be pleased.
Intel foresees the CXL bus enabling rack-level disaggregation of compute, memory, accelerators storage and network processors, with persistent memory on the CXL bus as well.
This was revealed when Intel presented a keynote pitch on the Compute Express Link (CXL) at the IEEE Hot Interconnects event.
CXL is based on the coming PCIe Gen-5.0 bus standard to interconnect processors and fast, low-latency peripheral devices.
Intel’s presenter was its Fellow and Director of I/O Technology and Standards, Dr Debendra Das Sharma and he started his session looking at Load-Store IO. This form of IO — loading and storing data into memory locations — is relevant because server memory capacity needs are rising due to the basic requirement to compute more data faster in AI, machine learning and other data-intensive applications such as genomics and big data analytics.
Load-Store IO is faster — much faster — than network IO, which transfers packets or frames of data, and is typically limited to taking place inside a server using CPU-level interconnect. Das Sharma said Load-Store IO physical layer (PHY) latencies are less than 10ns, whereas fast networking PHY latencies are in a >20 to >100ns range.
The aim is to extend Load-Store IO out of the server, and the way to do that is to use the PCIe Gen-5 bus as a base. The memory in connected devices can then be treated as cached, write-back memory by the server processor, and not need a DMA data transfer to move data between devices and the physical server CPU-attached memory. An Intel slide shows this:
Das Sharma mentioned three usage models for such CXL bus systems:
Type 1 — Caching Devices and Accelerators accessed via network interface cards (NIC);
Type 2 — Accelerators (GPU, FPGA) with their own memory, such as HBM;
Type 3 — Memory buffers for memory bandwidth and capacity expansion.
PCIe Gen-5 is fast enough to support server access to a pool of DDR5 DRAM across the CXL interconnect, and have it treated as usable DRAM by a server CPU. This, Sharma said, is poised to be an industry game-changer. It decouples compute from the traditional DIMM memory bandwidth and capacity limitations.
In the future NVDIMM (non-volatile DIMMs or persistent memory) could move to CXL, with DRAM backed up by storage-class memory (SCM) or NAND.
There could be computational storage devices — storage drives with on-board processors and memory and caching. They would do do on-drive compression, encryption, RAID, key:value store compaction, search or vector processing for AI/ML applications. They would have a DMA engine for moving data and use PCIe services such as NVM-Express.
Das Sharma says CXL enables systems to scale with heterogeneous processing and memory, with a shared cacheable memory space accessible to all using the same mechanisms.
In fact there could be a cluster-wide memory tier — a byte-addressable data store, scalable to petabytes:
CXL would provide a Load-Store IO fabric at rack-level, disaggregating compute, memory pools, accelerator pools (GPU, FPGA), storage pool and network processing units across racks.
A June 2021 version of Das Sharma’s presentation given at the EMEA Storage Developer Conference can be seen on YouTube.
Comment
The only SCM made by Intel is Optane. Das Sharma is talking about having Optane NVDIMMs hooked up to the CXL interconnect and have that capacity accessed by servers across the CXL link. He says persistent memory would then be cacheable, multi-headed for failover, and hot-pluggable.
As there is a 150+-member CXL consortium, we might expect other vendors’ SCM products to play nice in this space as well — such as potential ones from Micron, Samsung and others.
We think Das Sharma’s eventual rack-level disaggregation with CXL has a composability angle with sets of compute, memory, accelerator, storage and network processors dynamically composed to run specific application workloads.
