Home Blog Page 298

DDN ports EXAScaler storage software to Bluefield-2 DPU

By
Chris Mellor
-

DDN has demonstrated its EXAScaler storage controller software running on Nvidia’s BlueField-2 data processing units (DPUs) which replace the existing EXAScaler controllers and add a security layer to enhance data integrity.

EXAScaler is a Lustre parallel file system array. BlueField-2 and Bluefield-3 devices are host server offload accelerator systems looking after data centre networking, storage and cybersecurity capabilities.

DDN Marketing VP Kurt Kuckein wrote in a blog post yesterday: “EXAScaler Software Defined on BlueField eliminates the need for dedicated storage controllers and allows storage to scale on-demand and simply in proportion to the network.” Effectively DDN is supplying JBOD hardware and its EXAScaler software which runs in the Arm-powered BlueField-2 SoC.

This EXAScaler-BlueField integration was demonstrated at Nvidia’s virtual GPU Technology Conference on April 12.

Video of Oehme’s Nvidia GTC session.

Sven Oehme, DDN CTO, presented a session titled ‘The Secure Data Platform for AI’, during which he demonstrated an EXAScaler/BlueField-2 array linked to an Nvidia DGX GPU server and feeding it with data.

Oehme said that running the EXAScaler storage stack on BlueField meant less network management complexity, with BlueField managing everything above the base physical connectivity layer. BlueField also provided security measures that isolated the EXAScaler storage and its data from attackers. 

Kuckein wrote: “IO data now inherits all the benefits of BlueField’s network isolation, significantly reducing the number of attack vectors available to a malicious actor. Right away, hacked user accounts, rogue users, man in the middle attacks, hacked root accounts and other avenues for malicious activity are eliminated.” [His emphasis.]

The EXAScaler software stack was split in the demo between array (server) software running in the array’s BlueField-2 DPU and client (DGX) software running in the DGX GPU server’s BlueField-2 DPU. There was effectively a private 200Gbit/s network link between the two BlueField cards.

The demo showed multiple tenants, with each tenant seeing only their portion of the EXAScaler datastore and making virtual API calls to the EXAScaler array.  

Oehme said the initial software port to Bluefield-2 has been completed. Over the next couple of quarters, more DDN EXAScaler software releases will provide integration with server CPU bypass GPUDirect storage, end-to-end zero copy, and application Quality of Service.

The end-to-end zero copy means that data is transmitted from and to the array and the DGX with no time-consuming intermediate copying for data preparation at either end.

Comment

Here we have DDN’s EXAScaler array software running on proprietary Nvidia hardware. We don’t know how the performance of the BlueField-2 hardware (8 x Arm A72 cores) compares to the existing EXAScaler dual active:active Xeon controllers but suspect it’s a lot less.

An EXAScaler array using DDN’s SFA18KX hardware has four gen 2 Xeon Scalable Processors and delivers up to 3.2 million IOPS and transfers data at 90GB/sec.

We suspect that Bluefield-3, with 16 x Arm A78 cores, would be needed to approach this performance level.

When DDN and Nvidia add GPUDirect support to EXAScaler BlueField, that could provide DDN with a lower cost product than other storage suppliers supporting GPUDirect, such as Pavilion Data, WekaIO and VAST Data. The EXAScaler-BlueField-GPUDirect performance numbers will be very interesting to see.

Another angle to this DDN EXAScaler BlueField port lies in asking the question; which other storage suppliers could do likewise? DDN has said both NetApp and WekaIO are BlueField ecosystem members. Will either port their software to BlueField?

Lastly, Blocks & Files sees a parallel between DDN porting its EXAScaler storage software to a directly-connected and Arm-powered DPU and Nebulon developing its own Arm-powered Storage Processing Unit (SPU) hardware to provide storage array functionality using SSDs in the server hosting the Nebulon SPU card. The storage software stack runs in the DPU/SPU card and reads and writes data to drives in an attached chassis – JBOF with DDN, and server with Nebulon.

ExaGrid becomes cash-positive, passes 3,000 customers

By
Chris Mellor
-

ExaGrid recruited more than 120 new customers in the first 2021 quarter and has become cash-flow positive. The data backup startup said it also posted record sales in EMEA. “We have actually had three really strong quarters in a row. The company continues to grow,” ExaGrid president and CEO Bill Andrews told us.

The quarter saw ExaGrid bookings grew over 40 per cent and revenue 24 per cent as the total customer count moved past 3,000. Blocks & Files calculates that, since the third 2018 quarter ExaGrid has won an average of 60 new customers each quarter. “Remember, we go after the upper mid-market to the enterprise so these are not smaller customers,” Andrew said.

ExaGrid supplies a scale-out, globally deduplicating backup appliance with a so-called disk cache Landing Zone for fast restores of the most recent backups. There are two backup tiers; the Performance Tier for the next-fastest restores and the virtual air-gap Retention Tier for slower restores from deduplicated backup files.

Exagrid scale-out hardware.

The company in January announced v6.0 of its software, which contained the ransomware-combatting ExaGrid Retention Time-Lock feature with immutable backups that could not be deleted. Andrews said at the time: “ExaGrid has the only backup storage solution with a non-network-facing tier, delayed deletes, and immutable data objects.”’

Today, Andrews said: “ExaGrid has upgraded over 1,000 customers to Version 6.0 and all of those customers have the Retention Time-Lock for Ransomware Recovery feature turned on.”

The company claims a 75 per cent competitive win rate against suppliers such as Dell EMC (PowerProtect), HPE (StoreOnce),  Quantum (DXi), and Veritas.

Comment

Data protection startups are pushing the idea of backup-as-a-service for on-premises and in-cloud applications, provided by software running in the public cloud. The fastest restore for in-cloud apps will be from in-cloud backup stores, while the fastest on-premises restores will be from on-premises backup appliances.

That is ExaGrid’s market. It is a big market.

Dell EMC -provided IDC chart of Purpose-Built Backup Appliance supplier revenue market shares in 2020. Dell had 47 per cent of a market worth $4.33bn.

An IDC chart published this month by Dell EMC shows that the purpose-built backup appliance market generated $4.33bn revenues in 2020. Dell EMC soaked up 47 per cent share – to take approximately $2.035bn. With nothing else to go on but a rough visuals from the pie chart above, we think the unnamed second- and third-placed vendors look to have over 20 per cent market share. Who could they be? Our entirely unscientific hunches are that HPE StoreOnce is in second place. Is Exagrid the third-placed vendor?

Nvidia unveils BlueField 3 DPU. It’s much faster

By
Chris Mellor
-

Nvidia says its next-generation BlueField SmartNIC device, launched today, will deliver the most powerful software-defined networking, storage and cybersecurity acceleration capabilities for data centres.

Update: 13 April 2021. Updated Nvidia roadmap diagram added with revised SPECint numbers.

Doubled Arm core count, doubled network speed and PCIe Gen 5 support will make Nvidia’s third generation BlueField hardware deliver a five times higher SPECINT score than the gen 2 BlueField device. Nvidia is placing BlueField-3 in the AI and “accelerated computing” market, and says the DPU will free up x86 servers for application processing.

Jensen Huang, Nvidia founder and CEO, said in a press statement: “A new type of processor, designed to process data centre infrastructure software, is needed to offload and accelerate the tremendous compute load of virtualization, networking, storage, security and other cloud-native AI services. The time for BlueField DPU has come.”

Nvidia gained the BlueField technology via the Mellanox acquisition in March 2019. Bluefield is a system-on-a-Chip (SoC) which offloads a host server CPU by performing tasks such as interfacing to storage devices, security checking, and network data transmission. Sample shipping of BlueField-2 was announced in November last year and the device is now generally available.

BlueField-3 SoC.

BlueField-3 will have 16 x Arm A78 cores, providing ten times more compute power than BlueField 2, 400Gbit/s bandwidth and the aforementioned PCIe gen 5 support – four times faster than PCIe gen 3. Nvidia will include accelerators for software-defined storage, networking, security, streaming, line rate TLS/IPSEC cryptography, and – like BlueField-2 – precision timing for 5G telco and time synchronised data centres.

BlueField-3 is rated at 42 SPECint and 1.5 TOPS (TeraOps).

BlueField Roadmap. An earlier version of this roadmap depicted BlueField -2X and BlueField-3x devices each with an attached GPU. This is integrated into the BlueField SoC with the BlueField-4 generation.

In comparison BlueField-2 is equipped with eight x A72 Arm cores, 200Gbits/s Ethernet/InfiniBand networking, dedicated security engines and two VLIW accelerators. It supports NVMe over Fabric (NVMe-oF) Storage Direct, encryption, elastic storage, data integrity, compression, and deduplication. Bluefield-2 delivers 9 SPECint and 0.7 TOPS.

Nvidia’a DOCA software development kit (SDK) enables developers to create software-defined, cloud-native, DPU-accelerated services, leveraging industry-standard APIs, for BlueField-2 and BlueField-3.

Nvidia DOCA diagram.

Both BlueField products can isolate data centre security policies from the host CPU and build a zero-trust data centre domain at the edge of a server. BlueField-3 can also function as the monitoring, or telemetry, agent for Nvidia’s Morpheus cloud-native cybersecurity framework.“ Morpheus combines Mellanox in-server networking and Nvidia AI to do real-time, all-packet inspection to anticipate threats and eliminate them as they arise,” Huang said.

Morpheus uses machine learning to identify, capture and take action on threats and anomalies such as leaks of unencrypted sensitive data, phishing attacks and malware. It can analyse every network packet with line-rate speed and no data replication.

A BlueField ecosystem is being set up by Nvidia with:

  • Server suppliers Dell Technologies, Inspur, Lenovo and Supermicro integrating BlueField DPUs into their systems,
  • Cloud service providers Baidu, JD.com and UCloud using BlueField,
  • Support from system software suppliers Canonical, Red Hat and VMware,
  • Fortinet and Guardicore in the cybersecurity area supporting BlueField,
  • Support by storage providers DDN, NetApp and WekaIO, 
  • Edge IT platform support from Cloudflare, F5 and Juniper Networks.

BlueField-3 is backwards compatible with BlueField-2 and expected to sample in the first quarter of 2022.

Comment

BlueField is comparable to Fungible’s DPU in its scope but the Fungible Data Center processor and FS1600 Storage Processor are elements in Fungible’s composable datacenter scheme which uses its own network fabric. Nvidia is not presenting Bluefield as a means to the end of composing disaggregated data centre resources.

However, BlueField is software-defined. That implies that APIs could be developed so that servers and storage systems front-ended by BlueField could function in a composable data centre environment.

The diagram above showing BlueField roadmap to gen 4 is hardware-centric. Blocks & Files expects Nvidia to increase the software element as its ideas about the data centre future meet and react to composable data centre concepts being promoted by suppliers such as Fungible, HPE (Synergy) and Liqid.

We think Nvidia will in due course add Compute Express Link (CXL) support for Bluefield, as the CXL link can run across a PCIe Gen 5 bus.

Amazon Re:Freeze makes it easier to transfer Glacier vault archives

By
Chris Mellor
-

Amazon has announced the Glacier Re:Freeze serverless service to transfer an entire Glacier data vault to another S3 class such as Glacier Deep Archive.

The idea is to transfer an entire archive vault of data in a single service operation, with Amazon Lambda functions taking core of the component source vault inventory, restore, archive copy operations and data integrity checks. There is a dashboard to track progress.

An Amazon whats-new post states: “Deploying this solution allows you to seamlessly copy your S3 Glacier vault archives to more cost effective storage locations such as the Amazon S3 Glacier Deep Archive storage class.”

Glacier is Amazon’s S3 object storage service to store low-access rate data in a low-cost repository with data access taking minutes to hours. Glacier Deep Archive costs less than Glacier and takes longer to access data. Ordinary data access can take up to 12 hours with bulk data access at the PB level taking up to 48 hours.

Three points. Firstly, the copied Glacier vault is not deleted. A user has to do that manually. Secondly the Amazon post clearly states that Re:Freezer “automatically copies entire Amazon S3 Glacier vault archives to a defined destination Amazon Simple Storage Service (Amazon S3) bucket and S3 storage class.” It does not restrict the target S3 storage class to Glacier Deep Archive.

S3 Re:Freezer architecture diagram.

We take that to mean that the source Amazon Glacier archive contents can be copied to any S3 storage class: S3 Standard, S3 Intelligent Tiering, S3 Standard-Infrequent Access (IA), S3 One Zone-IA as well as S3 Glacier Deep Archive.

Our third point? We perceive potential nomenclatural confusion. If data is in the Glacier archive vault it is already ‘frozen’ and a service to transfer an archive to S3 Deep Archive ought to be called DeepFreezer – not Re:Freeze. Alternatively, a service to transfer it out of the Glacier archive to faster access S3 storage classes could be called DeFreezer. 

Blocks & Files envisages Azure, the Google Cloud Platform and other public cloud data archive services responding to this Amazon move with their own bulk archive copy services.

A Github ReadMe webpage contains a lot of information about Re:Freezer and deserves a look if you are interested in the idea of bulk copying of data from S3 Glacier to another S3 storage class.

Gartner: external storage spend to lag server spend 2019-2025

By
Chris Mellor
-

Server-based IT expenditure will grow three times faster than external storage spending, over 2019-2025, according to Gartner forecasts.

Wells Fargo senior analyst Aaron Rakers has presented tabulated Gartner data to his subscribers which reveals that total IT end-user spending will grow 5.9 per cent from $3.84tn in 2019 to $5tn in 2025. He noted: “Worldwide IT spending includes spending on end-user devices (PCs, smartphones, etc.), data centre systems (servers, storage, and networking), enterprise software, IT services, and communications services.”

Data Centre server spend will grow from $80bn in 2019 to £111.6bn in 2025, a 5.6 per cent CAGR. External controller-based storage was worth $28bn in 2019 and is forecast to reach $29.9bn in 2025, a 1.8 percent CAGR.

Table supplied by Aaron Rakers.

The big growth areas are in the software:

  • Enterprise Application Software; 12.4 per cent CAGR from 2019’s $217.8bn to 2025’s $380bn,
  • Infrastructure Software; 10.7 per cent CAGR from 2019’s $258.8bn to 2o25’s $422.8bn,
  • Managed Services and Cloud Infrastructure Services; 10.2 per cent CAGR from $472.7bn in 2019 to $692.4bn in 2025.

Wish HCI could do more?Maybe it’s time for some disaggregation

By
Joseph Martins
-
Lego bricks
Lego bricks

Sponsored Hyperconverged infrastructure (HCI) ticks a lot of boxes for a lot of organizations. Recent figures from IDC showed growth in the overall converged systems market was static in the fourth quarter of 2020, but sales growth for hyperconverged systems accelerated, up 7.4 percent and accounting for almost 55 per cent of the market.

It’s not hard to see why this might be. The pandemic means that many organizations are having to support largely remote workforces, meaning a surge of interest in systems that can support virtual desktop infrastructure. Those all-in-one systems seem to offer a straightforward way to scale up compute and storage to meet these challenges in a predictable, albeit slightly lumpy, way.

But what if your workloads are unpredictable? Are you sure that your storage capacity needs will always grow in lockstep with your compute needs? Looked at from this point of view, HCI may be a somewhat inflexible way to scale up your infrastructure, leaving you open to the possibility of paying for and managing storage and/or compute resources that you don’t actually need. Suddenly that tight level of integration is a source of irritation. Aggravation even.

This is why HPE has begun to offer customers a path to “disaggregation” with the HPE Nimble Storage dHCI line, which allows compute nodes, in the shape of HPE ProLiant servers, to share HPE Nimble storage arrays, while still offering the benefits of traditional HCI.

HPE’s Chuck Wood, Senior Product Marketing Manager, HPE Storage and Hyperconverged Infrastructure, says that while the classic HCI model delivers when it comes to deployment and management, admins still face complexity when it comes to the actual infrastructure.

“In traditional HCI, when you need to do Lifecycle Management, like adding nodes or even doing upgrades, all of those operations can be disruptive, because your apps and your data are on the same nodes,” he says.

VM-centric experience

At the same time, he explains, customers want to apply HCI to workloads beyond VDI and user computing: “There was this simplicity piece that you wanted to bring to more and more workloads – databases, business critical and mixed workloads.”

At first, it might seem counter-intuitive to split up the hardware elements that make up traditional HCI systems in the pursuit of simplification.

But as Wood explains, HPE’s approach is centered around delivering a simple VM-centric experience: “The VM admin wants to operate and manage workloads, virtual machines, you don’t want to be worrying about the infrastructure. To them, that’s where a lot of the complexity lies.”

So, in practice, “what we’re offering that’s aggregated is this simplified management … vCenter is the management plane, and our tools to manage Nimble Storage dHCI are embedded completely in vCenter.”

At the same time, Nimble dHCI has a Container Storage Interface (CSI) driver, so, as Wood puts it, “you can run VMs alongside containers. We can be that common infrastructure.”

“The disaggregated part is just the notion that you can add more ProLiants if all you need is memory, and compute and CPU resources,” he says. “Or you can add disks and all-flash Nimble shelves, if all you require is capacity for storage.”

This all allows for a much more incremental, and non-disruptive, approach to building out the infrastructure. For instance, adding HPE Nimble dHCI array to an existing setup of HPE ProLiants and approved switches is a five-step process that takes just 15 minutes, Wood says.

It’s worth noting that adopting HPE Nimble-based dHCI architecture does not require you to start from scratch. “If you have existing ProLiant Gen9 or Gen10 servers, we can actually bring them into Nimble Storage arrays,” says Wood, though to date, he says, most customers have built from the ground up.

Resilient infrastructure

Resiliency is further underpinned by HPE InfoSight, HPE’s AI-powered intelligent data platform, which aggregates telemetry from HPE Nimble systems to provide predictive management and automation. According to HPE, InfoSight can automatically resolve 86 percent of issues, while time spent managing problems is slashed by 85 percent.

The combination of ease of deployment and management, disaggregated architecture and automation via HPE InfoSight, means a commitment guaranteed to six nines of data availability for the HPE Nimble Storage dHCI systems. “It yields a better infrastructure from a resiliency perspective,” says Wood. “If you don’t have a lot of dynamics in your data center, and you just sort of set it and forget it for your general purpose workloads, then HCI is great, it’s perfect. But when it’s a more dynamic environment, where you don’t know what you’re going to need three to six months from now, disaggregated dHCI can be a better play.”

This inevitably prompts the question that, if you need to cater for changeable workloads, wouldn’t it make more sense to start using the cloud to supplement your existing storage infrastructure, before moving entirely off-prem.

Wood says the cloud is absolutely part of HPE’s offering in the shape of its HPE Cloud Volumes service. This gives customers the benefits of keeping control of their data on-prem, while having the ability to flex into the cloud as needed, further reducing the temptation to over-provisioning. HPE Nimble dHCI also offers support for Google Anthos.

And HPE Nimble dedupe and compression features further reduce the need for over-provisioning on prem, or in the cloud, he says. “Many customers are getting five-to-one dedupe on their backups and primary storage, so it gives you very good capability and resource utilization.”

On-prem

Even companies who began as cloud native are now finding they need to build out on-prem infrastructure, says Wood. “There’s a tipping point that ‘we have to have something on premises for test dev or to scale’,” he explains. Having on-prem infrastructure, supplemented by the cloud, gives customers far more control of their destiny, he says.

Wood cites the example of Australian pet insurance firm PetSure which had been using AWS to support VDI for its employees and provide remote veterinary services for customers’ pets. It switched to on-prem HPE Nimble dHCI as the Covid-19 pandemic hit, to support newly remote workers while also expanding its mixed workloads.

Similarly, he says, a healthcare customer in the UK opted to process X-rays locally, because it couldn’t afford to take the chance on latency that a cloud-based approach could involve.

“That’s why not everything is going to go to the cloud,” he says. “Because you need that local CPU memory and disk to give you the speed that your business requires to do your transactions.”

And, HPE would argue, you can now do that with a lot less aggravation, and aggregation, than you could in the past.

This article is sponsored by HPE.

Ionir frees data from its location trap

By
Chris Mellor
-

Profile Last month Blocks & Files interviewed Ionir marketing head Kirby Wadsworth and CEO Jacob Cherian to find out more about this storage startup.

Ionir stores block data using a generated name and virtual, time-sensitive location mapping metadata so that it can be moved in place and in time to enable persistent volume mobility – fast mobility to any point in time – for Kubernetes applications.

Ionir software provides persistent block storage volumes to cloud-native applications using Kubernetes. It uses technology developed by Reduxio, a company that failed to progress, and the former home of many Ionir founders.

Wadsworth said Ionir’s software is “elastic and dynamic with near-infinite capacity and performance,” and quite different from traditional block storage.

The location anchor

He said that traditional block storage data is defined by its location in a storage volume and an offset from the start of the volume. The volume is a logical location mapped to physical disk or NAND storage media.

Ionir’s technology accepts an incoming data write and generates a hash value from its contents, called its Name. It also generates a timestamp so that data with that name can be accessed by time.

Ionir data naming.

The data is then stored in a volume and has three basic metadata items associated with it; name, physical location (volume + offset), and time. The volume and offset are collectively called a Tag.

Now assume a second data write comes in and its generated hash has the same value as the first item. Then no new data is stored but new metadata calculated; the second item’s tag and time. This is intrinsic deduplication in the block datastore, “Any block repeating an existing one is mapped to it,” Wadworth said.

Blocks & Files combination of two Ionir slides showing data name, volume, offset and time relationships plus the inherent deduplication

If a third item comes in and its name is different, the new data is stored together with its metadata; name, tag and time. The chart above depicts this process, with unique data items represented by green lego bricks.

This combination of features means that the Ionir block store provides continuous data protection with no need for a separate data protection application. There is also no need either for snapshots or application quiescing to enable a snapshot to be taken.

The Ionir structure can be seen as a key:value store with the name being the key, and the data being the value. The data can be arbitrarily large, according to Ionir.

We suggested it was similar to object storage in that a data item’s name (Ionir) or address (object storage) is a hash generated from its contents. Cherian was emphatic in denying this: “It is not an object storage system.”

Data Teleport>

Here we have the basics for Ionir’s DataTeleport technology. If data has to be moved back in time – to a point just before a ransomware attack, for example, the restore to a previous point in time, using the timestamps, takes about a second; meaning a 1 second RPO.

If data has to be moved to a new location the metadata is moved first and this takes up to 40 seconds – it depends upon the size of the metadata store – after which the data can be accessed from the new location as the name in the metadata store is mapped to the original location. The location maps are altered as the data itself is physically moved, using smart pre-fetch algorithms. In effect the data is cloned.

Ionir says that its Data Teleport technology makes data portable, quickly portable in fact, and available, at any location in an on-premises and public cloud ‘space’ which can include multiple on-premises locations and public clouds in a multi-hybrid cloud infrastructure.

Cloud-native technology

Ionir’s software has three components, called the Pitcher, the Catcher and the Keeper.

The three Ionir software components.

The pitcher is the front end and receives data read and write requests from applications and compresses/decompresses data. Metadata is managed and accessed through the Catcher software and it carries out the intrinsic deduplication. Data is actually written to storage media by the Keeper component and it uses NVMe-accessed flash storage.

These three software components are written as microservices and are Kubernetes-orchestrated containers. Other containers request storage services from the Ionir containers using Kubernetes’ CSI interface.

The Ionir software supports automated tiering.

Its route to market is via a small direct sales force and channel partners such as the service providers outside the hyperscaler segment. Wadsworth said “Kubernetes levels the playing field,” and this benefits a small startup.

Future developments

What about the future?

Cherian said: Today, we store blocks of data, put together into volumes and presented to applications.” But: “A Pitcher could be developed to have a new interface.” An actual key:value interface was mentioned as an example in passing.

Also: “A new Keeper could be developed to support objects and different storage media. … We could have an S3 front end or backend. We’re storage protocol and access agnostic.”

Your occasional storage digest with Intel Ice Lake, composable systems and HCI

By
Chris Mellor
-

The big enterprise hardware event this week centres on Intel’s Ice Lake Xeon processors, with HPE saying its Synergy composable system will support Ice Lake compute nodes. We envisage a sequence of rolling upgrades across the industry’s HCI and external storage systems as suppliers adopt Ice Lake Technology.

Following on from last month’s launch of Fungible’s data centre processor, IDC has written a report forecasting the rise of composable, disaggregated systems, like Fungible’s.

Synergy with Ice Lake

HPE’s Synergy composable server will support Intel’s gen 3 Xeon Ice Lake CPUs, giving it a boost in processing power.

Synergy provides servers dynamically composed for workloads from pools of physical and virtual compute, storage, and network fabric resources. When the workload completes the components are returned to the pools for re-use.

The Synergy 480 Gen 10 plus is a 2-socket compute module with up to 40 per cent more than performance than gen 2 Xeon servers due to a faster CPU with up to 40 cores, 8 rather than 6 memory channels, use of the PCIe Gen 4 bus,  and support for Optane PMem 200 Series vs the earlier PMem 100 drives.

The 480 Gen 10 Plus will be available in the summer. 

Synergy systems are available as a service through HPE GreenLake.

Ice Lake, storage controllers and HCI

Blocks & Files expects hyperconverged infrastructure hardware suppliers and external storage system suppliers to announce powerful new products using Ice Lake gen 3 Xeon processors in the next few months.

Ice Lake CPUs have up to 40 cores, compared to gen 2 Xeon’s 26-core maximum and a third more memory channels than gen 2 Xeons; enabling 2TB of DRAM per CPU. These new processors also support the PCIe Gen 4 bus; twice the speed of PCIe Gen 3. 

Inspur says its M6 servers, which use Ice Lake, have overall storage density and IOPS (input/output operations per second) scaled up by 3 times and 3.2 times, respectively over its gen 2 Xeon servers.

All-in-all, this should enable at least 30-40 per cent more virtual machines supported in an HCI node, compared with today’s gen 2 Xeon boxes. HPE has confirmed to us that HPE SimpliVity HCI systems will support Ice Lake.

Intel Ice Lake Gen 3 Xeon CPU.

Object storage nodes using Ice Lake servers should get a similar boost in power as will scale-out filer nodes. External block arrays and unified block+file arrays should also receive a boost in power. We could expect raw capacity per controller to increase quite substantially and data reduction speed to accelerate as well.

Purpose-built back up appliances, such as Dell EMC’s PowerProtect line, should also be capable of significant performance and capacity Ice Lake boosts. Expect a continual sequence of Ice Lake updated storage and HCI systems over the next few quarters.

IDC composability report

IDC analysts have detected a divide that is opening up between current and next-generation workloads. They say composable, disaggregated systems can bridge the gap.

The IDC report ‘Composable/Disaggregated Infrastructure Systems: Catalysts for Digital Transformation’ explains that current workloads “assume infrastructure resiliency and require computing technologies such as virtualization and clustering that provide portability and transparently preserve application state. They also depend on external shared storage arrays that provide multiple layers of resiliency and redundancy.“

In contrast, next-generation and cloud-native workloads “are designed to be horizontally scalable, with state maintained in resilient data layer processes provided by the infrastructure itself. They are designed to run on industry-standard computing platforms and inside lightweight application containers and carry no bias toward specific silicon architecture.“

IDC’s report was produced by IDC Custom Solutions and is being distributed by Fungible, a maker of function-offload accelerators.

The report authors declare that “emerging composable/disaggregated infrastructure solutions provide the necessary scale, agility, and flexibility needed for cohosting current-generation and next-generation workloads on a common data centre infrastructure architecture.”

Fungible FS1600 Storage Server DPU

Composable/disaggregated infrastructure is based on “special silicon engines called function-offload accelerators (FAs) (also called data processing units or DPUs and SmartNICs) to disaggregate compute, storage, and networking resources into shared resource pools, which in turn can be allocated to workloads on demand in a software-defined manner.”

Shorts

Alluxio in collaboration with Intel has announced the availability of an in-memory software acceleration layer with 3rd Gen Intel Xeon Ice Lake CPUs Optane persistent memory (PMem) 200 series drives. It’s aimed at analytics and AI pipeline workloads and uses Optane’s Storage over App Direct feature.

Research outfit Counterpoint says average smartphone NAND flash capacity crossed the 100GB mark in 2020 for the first time. It was 140.9GB in IOS phones in the fourth 2020 quarter but 95.7GB in Android phones during the same period.

Google Database Migration Service is available to support enterprises migrate their databases and other business IT infrastructure to Google’s public cloud. Freedom Financial Services used it and the Google Kubernetes Engine (GKE) to migrate its 1TB MySCQL database to Google’s Cloud SQL in 13 hours.

M-Files, an information management company, today announced the acquisition of Hubshare and its secure information exchange platform to bolster external content file sharing and collaboration.

Pavilion Data Systems is partnering with Graphistry to supply data analytics functionality using Nvidia Magnum IO and GPUDirect technologies. Graphistry has integrated data UI tools like Jupyter notebooks, Streamlit dashboards, and Graphistry visual analytics with the Nvidia RAPIDS GPU ecosystem (Dask_cuDF, GPU Direct storage, DGX A100) and with the Pavilion HyperParallel Data Platform.

Platform9 has updated its Platform9 Managed Kubernetes(PMK) product with IPv6 support for 5G deployments. It’s claimed to deliver near line-rate performance for network packets.

SaaS business Redstor has announced new Azure services. There is an AI engine to detect malware, and the ability to back up and recover Kubernetes environments on Azure. Redstor can also migrate and recover full systems to Azure from any cloud or data centre. 

ScaleOut Software has released an on-premises version of its ScaleOut Digital Twin Streaming Service with all of the analytics and visualisation tools available from its Azure-hosted cloud version. It enables customers to organisations develop and deploy streaming analytics applications on their own servers and so obviate public cloud security concerns.

Cloud storage supplier Wasabi is partnering with Atempo to offer the Tina backup and restore functionality across physical and virtual servers for applications, databases, NAS, Office 365, and more.

The Linac Coherent Light Source (LCLS) at the Stanford Linear Accelerator Centre (SLAC) National Accelerator Laboratory has deployed WekaIO filesystem software to store data for its LCLS Free Electron Laser (FEL) X-ray snapshots of atoms and molecules. These provide atomic resolution detail on ultrafast timescales to reveal fundamental processes in materials, technology, and living things.

Storage product comparisons on the rack

By
Chris Mellor
-

Analysis. Comparing the performance of storage products is bedevilled by the lack of benchmark testing for the three main storage storage access methods; blocks, files and objects. There are industry standard benchmarks but few suppliers use them and comparing results when a benchmark has five or more aspects can be tremendously difficult.

Blocks & Files recently explored this topic when comparing Nvidia DGX A100 GPU server storage vendors and found it a complete pain. There was no straightforward way to compare the storage vendors because they – and Nvidia – tested different storage configurations with varying hardware characteristics and varying GPU server target configurations.

We tried generating a single storage node to single DGX A-100 throughput measure across the suppliers. This ran into the problem of defining a storage node. Is a NetApp A-800 high-availability pair ‘node’ equivalent to a Pavilion HyperParallel array ’node’ and are they equivalent to a VAST Data Universal Storage array or a WekaIO cluster all of equivalent capacity?

Their physical size and performance vary widely. 

Normalising factor

We decided it was obvious that there was no single normalising factor; not throughput, not IOPS, not rack units and not product SKU. Instead we realised that we needed a base factor to which other things could be related. This base factor had to be common across products and represent something that was relevant to enterprise customers.

What ever storage product they have, be it software sold on its own or a HW+SW combination, has to fit in a standard rack and deliver data to accessing systems in block, file or object format, or some combination of those

Data centre racks.

Rack unit normalisation

We realised that rack units could be our base normalising factor.

Whatever storage product is used it fits in a rack and its performance on any basic protocol; block file or object, can be expressed in units per rack unit or U for short.

For example, let’s envisage a HW+SW filer. It could be a 5U filer delivering 750,000 read or write IOPS, 50MB/sec read bandwidth, 35MB/sec write bandwidth, and costing $350,000. These IOPS, bandwidth and cost numbers can be divided by 5 to produce IOPS/U, read and write MB/U and cost/U. The resulting numbers are 150,000 read/write IOPS, 10MB/sec read, 7MB/sec write and $70,000.

This normalised-to-U approach provides a basis for comparing the 5U HW+SW filer with a SW-only file system product like WekaIO or VAST Data’s Universal Storage platform.

Envisaging numbers

The WekaIO system runs in server cluster nodes. Let’s envisage an imaginary  3-server WekaIO config using three 2U servers which outputs 1 million read or write IOPS, 75MB/sec read, 50MB/sec write and an overall cost of $400,000. We divide these numbers by 6 to derive per-U numbers, such as 166,666 read/write IOPS, 12.5MB/sec read and 8.3MB/sec write bandwidth, and $66,667.

Immediately we can compare the 5Y HW+SW filer with the WekaIO system and see the WekaIO system is cheaper; $66.667/U vs $70,000/U. Similarly it has better read/write IOPS and read and write bandwidth than the 5U filer in our fantasy comparison.

All these numbers are made up and exist only to illustrate the approach. You use the attributes that are valid to you, such as subscription cost over three years instead of acquisition cost plus three years support. And then you compare various candidate products and normalise their values on these attributes to per-U numbers.

Worked example

We decided to try a worked example and looked at block storage and started building a spreadsheet using whatever publicly available supplier numbers we could find;

The product numbers are on the left side of the table with the per-rack unit numbers in the grey stippled area to the right. There are a fair few missing numbers, as not all array suppliers publicly reveal their performance numbers.

We can chart columns of interest to make supplier comparisons easier. For example, raw capacity per rack unit:

We recognised that our spreadsheet could have cost and latency numbers added but also saw that supplier data sheets miss out many numbers, such as random write IOPS, random read and write latencies, cost and so forth.

If customers were to use this approach of normalising product attributes to rack units they could ask their candidate vendors for these numbers.

Conclusion

This normalisation to rack unit approach seems to provide a valid way to compare multiple different suppliers’ products on basic performance and cost attributes, with a couple of caveats:

  1. Performance tests should be run with multiple clients to fully saturate the hardware/software system.
  2. The same client and network configuration should be used for each supplier product test.

It’s applicable to on-premises block, file and object storage systems, be they external shared storage, software-defined storage or hyper-converged systems. It could not be used to compare public cloud storage offerings though. That would need a different approach, one that our brain cells cannot yet work out.

Ice Lake scales MemVerge in-memory SW to new heights

By
Chris Mellor
-

Updated Memory Machine software from MemVerge runs applications faster with support for Ice Lake CPUs and increased memory capacity.

MemVerge CEO Charles Fan issued this announcement: “Memory Machine v1.2 is designed to allow application vendors and end-users to take full advantage of Intel’s latest Xeon Scalable processor and Optane memory technology.”

MemVerge’s Big Memory software virtualizes DRAM and Optane persistent memory (PMem) tiers into a single resource pool to host applications and their working data set in memory and so avoid making time-sapping storage IO calls to SSDs or disk drives. 

Blocks & Files diagram showing Memory Machine concept.

Memory Machine 1.2 supports four to 80 Ice Lake cores and up to 6TB of DRAM + Optane PMem 200 storage-class memory per Ice Lake CPU. There is 32 per cent more bandwidth than with Optane PMem 200 than PMem 100 drives.

Newly-announced Ice Lake gen 3 Xeon CPUs run faster than gen 2 Xeons which speeds up in-memory apps. Ice Lake also supports more memory sockets – 8 rather than 6 – giving 2TB DRAM capacity per CPU, instead of the prior 1.5TB. Ice Lake Gen 2’s Optane PMem support bulks up the overall per-CPU memory capacity to 6TB from gen 2 Xeon’s 4.5TB maximum. This extra capacity means more and larger in-memory applications can run and execute faster.

Storage Review testing showed that Memory Machine 1.2 running with dual 40-core Ice Lake CPU, 512GB DRAM, and 2TB of Optane PMem 200 provided 2x read performance and 3x write performance of a dual 26-core gen 2 Cascade Lake Xeon system with 192GB DRAM and 1.5TB of PMem 100 capacity.

MemVerge’s press announcement includes a quote from Mark Wright, technology manager for Chapeau Studios: “Initially, we opened a poly-dense scene in Maya and it took two-and-a-half minutes [from storage]. Then, we opened a scene from a snapshot we’d taken with Memory Machine and it took eight seconds.”

V1.2 Memory Machine adds:

  • Centralised Memory Management for configuration, monitoring, and alerts for DRAM and PMem across the data centre,
  • Redis and Hazelcast Cluster high-availability through coordinated in-memory snapshots to enable instant recovery of the entire cluster,
  • Double the OLPT performance of Microsoft SQL Server on Linux ,
  • Support for QEMU-KVM hypervisor with dynamic tuning of DRAM:PMem ratio per VM, and minimised performance degradation caused by noisy neighbours,  
  • Autosaving and in-memory snapshots allow animation and VFX apps to provide Time Machine capabilities that allow artists to share workspaces instantly and recover from crashes in seconds.

MemVerge expects the software upgrade will accelerate single-cell genome analytics but has not yet published figures demonstrating this.

The company has joined the CXL consortium, which is developing coherent bus technology to enable remote access to pools of memory. The company has also set up labs at Arrow, Intel, MemVerge, Penguin Computing, and WWT that are equipped for Big Memory demonstrations, proof-of-concept testing, and software integration.

MinIO adds mainstream object storage management features

By
Chris Mellor
-

MinIO, the Kubernetes object storage developer, this week released three management tools that collectively “allow customers to rapidly adopt even the most advanced MinIO capabilities while providing a roadmap for increased automation as they scale their deployments and workloads.”

MinIO Kubernetes Operator enables enterprise admins to deploy and manage object storage independently of the underlying edge, data centre and public cloud infrastructure. MinIO Console is a facility to provision multi-tenant object storage as a service. SUBNET Health runs automated diagnostic checks across storage drives, the network, CPU, memory, OS, containers and MinIO software components.

These three tools bring MinIO open source object storage closer to mainstream enterprise object storage products, such as those from Cloudian, Scality, NetApp (StorageGRID) and others.

In more detail, Operator includes DevOps tasks such as tenant creation, expansion and management capabilities for enterprise admin staff. It is available on major Kubernetes distributions including Red Hat OpenShift, VMware vSphere 7.0u1, SUSE Rancher, and HPE Ezmeral, and public cloud providers; Amazon’s EKS (Elastic Kubernetes Engine), Azure’s AKS (Azure Kubernetes Service), Google’s GKE (Google Kubernetes Engine) and Anthos.

MinIO Console and its GUI enables cluster management and help admin staff check system health, perform audit tasks and integrate MinIO software with other components. It’s built to support large scale deployments.

SUBNET Health is a “critical component of MinIO’s hybrid cloud strategy and provides customers with the confidence to run on any hardware, in any environment and on any cloud,” MinIO CEO and Co-Founder AB Periasamy said in a statement

Operator, Console and SUBNET Health are licensed under the GNU AGPL v3 license. An alternative commercial license is priced on usable capacity and users get access to the MinIO Subscription Network (SUBNET) with 24/7/365 direct-to-engineer support, access to the Panic Button, SUBNET Health, architecture, performance and security reviews.

DDN speeds up AI storage system in a flash

By
Chris Mellor
-

DDN has upgraded its A³I artificial intelligence -dedicated storage system, using flash to speed data access, as well as improving system monitoring tools.

The systems now support all Nvidia DGX servers, including the DGX A100 and SuperPod, with reference architectures for both. DDN launched the A³I in October 2018 as a turnkey, AI-dedicated storage and Nvidia DGX-1 GPU server product set, featuring all-NVMe flash and hybrid flash/disk Exascaler arrays and Lustre parallel-access file system software.

Dr. James Coomer, DDN’s VP products, said in a statement: “With huge performance boosts and innovation in operational efficiency, this new, differentiating feature set eliminates complexities and strengthens our AI supercomputing infrastructure, making it easier to manage for data-intensive scalers and data centric enterprises, alike.” 

DDN A³I system

The monitoring improvements include:

  • End-to-end performance monitoring for DGX and DGX pods with compute, storage and network status visibility
  • New GUI
  • Storage quota monitoring,
  • Workload and system status
  • Call home feature

Performance boosting

The “huge” performance boosts aren’t quantified but are based on ‘Hot Nodes’ caches and ‘Hot Pools’ tiered flash and disk.

The Exascaler systems come with ‘Stratagem, an integrated policy engine that moves data automatically between disk and flash storage tiers. This means active data faster is stored in flash, thus enabling faster access. With this week’s upgrade, Stratagem’s data placement algorithms have been improved to better determined how hot files are based on recent access. Overall, this lowers data access latency and response times, DDN said. There are also new API

Hot Noding

We asked DDN questions about the Hot Nodes.

Blocks & Files: What is a Hot Node?

DDN: A Hot node is a Filesystem client attached to the DDN shared parallel storage appliances, with local NVMe devices that can be used as actively managed buffer between shared and local data. DDN’s A³I filesystem manages the population of the Hot Node local storage to reduce unnecessary access to shared storage, and to reduce latencies for access to the hot data.

Blocks & Files: How much flash storage is used?

DDN: All that’s available. With a single NVIDIA DGX we have around 15TB of local NVMe [flash] in each system.

Blocks & Files: How does giving a hot node some flash storage remove the risk of manual data movement?

DDN: Intelligence in software works to automatically ensure that data is positioned at appropriate time when needed. For most common current read-only-from-cache AI use cases, entire data sets are brought from shared to local space as it’s being read with no detriment to performance. Tested with up to 560 nodes simultaneously and working fine.

Blocks & Files: How do the Hot Nodes get “Excellent IO performance?

DDN: In early (pre beta) at scale testing 10X+ acceleration with Hot Nodes when running IO at scale.

Blocks & Files: How are they different from non-Hot Nodes?

DDN: Hot Nodes are those Compute clients which are designated to have a managed local cache. Typically, hot nodes have accelerator (GPU, CPU or other), high performance network connection to storage and local NVMe devices to act as managed data cache. Non-hot nodes don’t have local managed data cache. It’s a switch in DDN’s A³I filesystem that determines whether a node is hot or not.

Note; DDN did not provide actual performance numbers.

1...297298299...438Page 298 of 438

ABOUT US

Blocks & Files is a storage news, information and analysis site covering storage media, devices from drives through arrays to server-based storage, cloud storage, networking and protocols, data management, suppliers and standards. We publish news, views, opinions and analysis of storage industry events and developments, aiming to keep storage users, developers, consultants, marketeers, salespeople, support staff, analysts and everyone else involved in IT storage up to date with what’s going on.

FOLLOW US

© Situation Publishing, 2018-2024