Data Centre

Storage

The solid state of storage in 2018: Latencies, they are dwindling. On-premises, the kit is glistening...

A beautiful sight, we're happy tonight... Walking in a storage wonderland

By Chris Mellor

11 SHARE

Roundup 2018 wasn't too shabby for storage as the three-pronged attack of IoT edge, AI and multi-cloud defeated notions of the public cloud eating on-premises storage. Revenues also rose, capacities increased and latencies dived to boot.

Capacity rises followed new technology promising extra layers and QLC (4 bits/cell) for flash, and assisted recording techniques for disk.

Storage-class memory stepped closer, NVMe-over-Fabrics arrived and composable systems gathered strength. Hyperconverged systems grew their market presence as well.

There were 19 acquisitions, 22 company funding rounds and only two exits, both followed by rebirths. The big four – Dell EMC, HPE, IBM and NetApp – did well and many startups prospered, but one or two companies went through restructurings.

Let's start with the public cloud pushback first, and then look at media advances, take a canter around arrays and filesystems, and then acquisitions, funding, restructurings and exits.

Public cloud threat diminished

Amazon, Azure and Google are the three killer whales threatening to decimate on-premises storage. But the rise of IoT edge, where the public cloud meets IoT data-generating devices, soon put a stop to that.

Edge devices generate a heck of a lot of data. Think aeroplanes, ships and autonomous cars with sensors detecting hundreds of thousands of events per hour, which may need an immediate response. There's too much data to send to a public cloud because of network bandwidth limitations, and it takes too long for the public cloud to receive, analyse, respond and send it back to the edge device.

In effect, small and complex edge devices need their own micro-data centres and storage. We can look at data-generating branch offices and define them as "internet edge locations" as well. Superstore branches are edge locations and need their own on-premises IT.

Machine learning and AI technology is also needed at the edge, where stored, locally generated data can be analysed in real time for local decision-making. We can send filtered results to the public cloud for central analysis, archiving, etc, and so have a nice partnership between the on-premises and public cloud worlds.

A third factor is multi-cloud, a combination of public and private (on-premises) clouds, where customers avoid lock-in by using two or more vendors, often coordinated though a central on-premises hub. Tellingly AWS realised it needed an on-premises presence for more than data collection, which its Snowball devices provided, and Outpost was the result.

The release of Amazon's Outpost and Azure Stack-like devices in November was a watershed moment. Now even Amazon supports the multi-cloud idea.

Overall, many data centres have become internet edge data centres, meaning on-premises storage has a future.

Non-volatile media

There was activity on three non-volatile fronts: layering, quad-level cells (QLC or 4bits/cell) and storage-class memory.

SSDs using 64-layer flash, organised into TLC (3 bits/cell), predominated and all the manufacturers announced plans to move to 96-layer flash, 90+ layers in Samsung's case, and by so doing, add 50 per cent more capacity.

China's YMTC NAND foundry startup said it intended to jump straight from its 64-layer flash to 128, bypassing the 96-layer stage altogether. This would enable it to catch up with Intel, Micron, SK Hynix, Toshiba and Western Digital as they eventually transition from 96 to 128-layer NAND.

Controller and NAND technology developed sufficiently to support quad-level cell (QLC) technology with 4 bits/cell, a third more than TLC. QLC has lower write endurance than TLC and is thus more suitable for read-intensive work, such as fast-access archiving.

QLC SSD announcements came from Micron in February with its 5210 ION, from Western Digital with its 96-layer flash in July, and Samsung and Intel rolled it out in August.

SSDs using QLC flash could eat into nearline disk drive sales, but capacity improvements coming for disk should preserve a price advantage for the three disk drive manufacturers.

There were also heroic SSD capacity demonstrations, such as Samsung's 31TB drive in February.

Non-events and uncouplings

What didn't happen with flash in 2018 was the adoption of high-capacity, ruler-format SSDs by server suppliers. Samsung's faster version of NAND, its Z-SSD, also seemed to splutter.

The two big events in the flash industry were, firstly, Micron and Intel deciding to dismantle their joint IMFT foundry business and go their separate ways in manufacturing and developing both 3D NAND and 3D XPoint.

Secondly, Toshiba sold its Toshiba Memory business for $18bn to a Bain Capital-led consortium. The consortium owns Toshiba's interest in its flash foundry joint venture with Western Digital. Seagate was a consortium member and now has a guaranteed supply of chips.

Storage-Class memory

Intel steadfastly carried on pushing Optane into the market, with Cisco and Lenovo aadopting it for some servers. Liqid added Optane SSDs to its composable systems server component mix late in December.

Israeli firm Scale MP has software treating Optane SSDs as storage-class memory and both Intel and Western Digital OEM that software.

HPE upgraded its 3PAR array operating system to use Optane SSD storage-class memory caching, with Nimble array support coming. Dell EMC will add Optane storage-class memory support next year.

E8's NVMe-oF software supports Optane.

NetApp announced its Optane-supporting MAX Data, with server Optane being used to cache/tier data fed it by a NetApp Array.

The main product event was Intel's Optane DIMM announcement in August. These provide a 350-nanosecond average read latency, compared to the Optane SSD's 10,000. This should help increase Optane adoption by server OEMs.

Slowly but surely, a wave of SCM adoption is building and we should see greater adoption next year, with radically faster data access by servers as a result. Let's hope that Samsung does something with its Z-SSD to provide competition.

The disk world

Disk capacities climbed during the year, with a focus on 3.5-inch nearline drives. These grew from 10 through 12 to 14TB with all three manufacturers providing a range of drives for desktop, data centre, NAS and surveillance use.

Western Digital even produced a 15TB shingled drive, but the next capacity level for mainstream drives will be, we understand, 16TB. This will use, and be the last gasp of, perpendicular magnetic recording.

Higher capacities mean smaller bits, and the only way of keeping their magnetic polarity stable is by using different recording mediums which need heat or microwave radiation given off by disk read/write heads to change polarity.

Seagate is using Heat-Assisted Magnetic Recording (HAMR) to do this, and is testing 16TB HAMR drives. Western Digital and Toshiba are using Microwave-Assisted Magnetic Recording (MAMR) and should have drives to test next year, if they don't already.

We should see MAMR and HAMR drives arriving in 2019 with capacities at the 16-20TB level.

During 2018, Seagate revealed it will introduce a 14TB drive with multi-actuator technology next year. This involves the use of two read/write heads per platter to increase a disk drive's IO rate.

We have seen behind-the-scenes technology development this year which should deliver 20TB-plus drives next year, as well as drives with higher IO rates. Both will be welcome.

We also saw manufacturing footprint retrenchment, as SSDs ate into notebook, desktop, 2.5-inch and fast 3.5-inch data centre drive sales. For example, Western Digital closed a Kuala Lumpur facility in July.

Tape lives on

IBM launched is TS1160 tape drive and format, supporting 20TB raw capacity, compared to LTO-8's 12TB. But LTO-9 with 24TB capacity is expected next year.

Spectra Logic added the TS1160 drives to its TFinity Exascale library, claiming capacities greater than 2EB were now possible.

Great stuff, were it not for Amazon announcement in November of its Glacier Deep Archive cloud storage, priced at $0.01/TB/month. The effects of this on the on-premises tape library business could be significant.

Disk and flash arrays

Disk arrays grew capacity and speed during the year. Dell EMC significantly upgraded the speed of its Unity and SC arrays.

There were two other themes to array development: NVMe-oF and analytics.

Many arrays already support NVMe SSDs. Startups offer NVME-oF access which provides near PCIe direct-attached storage speed across a network link to an external array; think E8, Excelero and Pavilion.

NVME-oF has depended upon data centre-class (lossless) Ethernet or InfiniBand for the wire transport. But NVMe-over-TCP arrived, as did NVMe-over-Fibre Channel, and both provide a way for installed base SANs to upgrade to NVMe-oF without adding InfiniBand or lossless Ethernet.

NetApp announced support for Fibre Channel NVMe. IBM has end-to-end NVMe on its Flash System arrays. Pure Storage is adding NVMe-oF support, as is HPE.

Basically, the NVMe-oF groundwork has been done and mainstream adoption will start next year.

The second theme is that of predictive analytics, from a cloud centre, being used to optimize the operation of on-premises arrays, Nimble showed the way with its InfoSight technology. HPE bought Nimble and has now spread InfoSight to its 3PAR arrays.

Pure Storage has added predictive analytics. Hyperconverged system vendors are adding it. Soon it will be table stakes for any enterprise storage array vendor.

New array suppliers

Flash chip and disk drive suppliers Toshiba and Western Digital extended or entered the drive array business. Toshiba's Kumoscale software runs in a server front-ending a set of all-flash drive array shelves, such as ones from Lenovo. It virtualizes and abstracts the SSDs, turning them, we were told, into one exceptionally large SSD.

Western Digital is more mainstream. Its Ultrastar Serv60+8 is a hybrid 4U enclosure with two Xeon controllers, 60 x 3.5-inch disk drive bays and 8 x 2.5-inch SSD bays. It can hold up to 780TB of capacity with 12TB helium-filled disk drives, passing 900TB once 14TB disks and 7.68TB SSDs become available.

The company also introduced two secondary or archive disk storage arrays, ActiveScale P100 and X100, with the v5 ActiveScale OS based on HGST-acquired Amplidata object storage.

Western Digital ActiveScale racks

Whether Western Digital will become a significant and serious array supplier, as well as a composable systems supplier, remains to be seen. Selling arrays and systems is a lot different from selling drives to OEMs, VARs and consumers.

Benchmarks

There was a lot of filer benchmarking with SPEC SFS 2014. This saw seven results for its Software Build test added in the year:

Vendor Builds Overall Response Time Date
NetApp A800 8-node 4,200 0.78 Nov 2018
NetApp A800 4-node 2,200 0.73 Nov 2018
DDN - Spectrum Scale 1,500 0.19 Oct 2018
E8 - Optane - Spectrum Scale 1,200 0.57 Aug 2018
WekaIO - Supermicro 1,200 1.02 Mar 2018
Huawei 6800F V5 1,000 0.59 May 2018
E8 - Spectrum Scale 600 0.69 Jan 2018

There had been three results added in 2017 and two in 2016; so a marked increase was seen in 2018, as well as a sevenfold increase in speed.

The ever-popular SPC-1 (v3.n), which tests storage array responsiveness, had 14 results added in the year, with 10 from Chinese or other Asian suppliers. The top SPC result in 2017 was a Huawei FusionStorage system scoring 4,500,392 IOPS, so the top score has increased substantially.

Vendor SPC-1 IOPS Price/performance ($/KIOPS) Date
Huawei 18000 v3 7,000,565 $376.96 Oct 2018
Huawei 18800F V5 6,000,572 $465.79 Mar 2018
Huawei 18500F V5 4,800,419 $400.96 May 2018
JetSpeed 10-node 2,410,271 $287.01 Dec 2018
NetApp A800 2,401,171 $1,154.53 Jul 2018
JetSpeed 7-node 1,510,150 $326.75 Dec 2018
Inspur AS5300G2 1,500,346 $307.62 Oct 2018
Fujitsu AF650 S2 620,163 $269.79 Jan 2018
Fujitsu DX600 S4 560,037 $295.68 Apr 2018
Lenovo DE6000H 460,011 $91.76 Nov 2018
Huawei 5300 V5 450,212 $416.27 Jul 2018
Fujitsu AF250 S2 360,070 $94.02 Nov 2018
MacroSAN MS2500G2 - 25E 300,028 $379.50 Feb 2018
Lenovo DS6200 180,006 $93.29 May 2018

The only American supplier represented is NetApp. In effect US suppliers are abandoning the SPC-1 benchmark.

The rise of converged AI systems, combining storage and Nvidia DGX-1 GPU servers saw a series of new benchmarks appear, such as Resnet-152 and Resnet-50 image recognition for machine learning.

There were results from Dell EMC (Resnet-50), DDN, IBM, NetApp and Pure Storage released throughout the year. Here's a sample chart:

STAC benchmarks also appeared, tailored specifically for financial trading and similar finance applications. These are tightly controlled by a user panel and divided into specific versions for individual financial trading workloads, making them hard to use for general storage performance purposes.

Dell, E8 and Lenovo have participated in some of these STAC runs.

Filesystems

IBM's Spectrum Scale notched up sales and benchmark wins while newer startups such as Elastifile, and WekaIO pushed their products. WekaIO is going for Spectrum Scale levels of performance with greater ease of use. Elastifile is relying on the multi-cloud idea while still being scale-out.

HPC supplier Panasas announced new and faster software (PanFS) and changed tack from proprietary to commodity hardware, preparing the way for a possible software-only release or cloud instantiations of PanFS.

Qumulo added more speed to its filers and software. Google announced its Cloud File Store. Basically, file storage boomed in 2018.

Object storage

But it was a fairly quiet object storage year. Suppliers continued adding S3 interfaces and file-level access. Cloudian and Scality received fresh funding and increased their mainstream storage supplier partnerships.

Small Caringo put out an impressive Swarm software release, claiming an S3 throughput record. Minio recorded record GitHub pulls for its software.

We saw the beginnings of all-flash object storage nodes, promising much faster performance, but we also saw new filer companies, like WekaIO, promising object-storage levels of scale-out with filer performance.

It's almost as though object storage is a mature market.

Backup and data managers

The huge growth of Veeam, followed by Acronis, Cohesity, Rubrik and others, caused changes that caught legacy suppliers Commvault and Veritas on the hop.

Backup became seen as a generator of secondary storage; unstructured data files that could be managed to do more than offer data protection. Secondary storage is much more than backup and several suppliers have not shown interest in progressing beyond data protection into data management. Veeam, for example.

However, every secondary storage company is now in the data management business, including Actifio, Cohesity, CTERA, Igneous, Komprise and Rubrik. Even Scality has its Zenko data manager.

The backup companies are slowly being transformed into data managers; witness Commvault and Druva.

All these data managers rely on building metadata warehouses describing the unstructured data files they look after, and then using this to deliver data copies for test-and-dev, and provide a mix of archiving, analytics, data protection, security, governance and life cycle management.

Backup suppliers benefited from an increasing willingness by customers to work with multiple vendors, choosing best-of-breed products for specific backup workloads.

There was a smattering of consolidation, with Carbonite buying Mozy, and Kaseya buying Unitrends and Spanning.

Acquisitions

We counted 19 acquisitions in 2018:

IBM buying Red Hat for $34bn had a storage element but it wasn't about storage per se.

Most of the acquisitions above could be called tuck-ins, with technology being bought that helped the acquirer strengthen its market presence, like Nutanix and Netsil, or enter a nearby adjacent market, like Rubrik and DatosIO.

The Bain group purchase of Toshiba's Memory Business was atypical, being a private equity rescue bid for a business Toshiba had to offload to recover from drastic financial losses in its nuclear power station building business.

Exits and arrivals

There were remarkably few storage company failures in 2018; just two notable in fact, one minor, and also one notable arrival.

Primary Data closed down due to a focus on on-premises storage restricting its sales growth. However, CEO and founder David Flynn popped up again later in the year, in the same building and with many Primary Data staff, with Hammerspace. It is a metadata-based entity, like Primary Data, but focused on unifying file silos into a single network-attached storage (NAS) resource that can provide access to unstructured data anywhere, in hybrid or public clouds, on demand.

All-flash and hybrid array supplier Tintri went belly up after a catastrophic IPO. In September its assets were bought for $60m, a surprisingly large sum, by HPC storage supplier DDN, which is entering the enterprise storage array market by renovating the business.

It has just announced the activation and stocking of more than 98 per cent of the planned 122 Tintri service depots worldwide, and the general availability of Tintri OS 4.4.3.3, the first OS release since DDN acquired it.

Data Torrent, a startup focused on the Apex stream processing engine, closed down in May, having failed to attract sufficient customer interest.

Restructurings, CEO changes, etc.

Three companies are undergoing restructuring to mend failing or sub-optimal business models: Commvault, Quantum and Veritas.

Commvault is changing its cost structures and business model with product simplification, a renewed channel push and a move to subscription pricing.

Quantum found its StorNext scale-out file system management software wasn't generating the revenues needed to offset tape and deduplicating disk array products.

Jamie Lerner was appointed CEO in June 2018, the fourth CEO in eight months. This was triggered by CEO Jon Gacek's departure in November 2017, following the intervention of activist investor VIEX and a March 2017 NYSE delisting threat sidestepped a month later by a reverse stock split.

The company is also investigating accounting irregularities and was late with a quarterly SEC results filing. Lerner has instituted a vertical market strategy with the StorNext product set integrated with workflow stages and applications in markets such as entertainment and media, oil and gas and automotive.

Private equity-owned Veritas is working to cut costs and become profitable. There have been reorgs and layoffs. Legacy products and some new ones are being examined for viability.

Other, smaller companies saw CEO changes – Datrium, Elastifile, Pavilion, Reduxio and Sysdig, for example – as they aimed to kickstart growth or accelerate it.

Veeam's October CEO change was not associated with any dip in growth or revenues at this rocket-ship company. Peter McKay left to pursue new endeavours.

Funding

We recorded 22 storage company funding events in the year:

That totals more than $2.6bn and indicates just how strong the storage industry is.

The rise of the AI-driven analytics trend, the potentially colossal data-generating capabilities of internet edge devices, and the deferred or weakened gobble of on-premises storage by the public cloud suggests that the storage industry is in good health and should be even stronger this time next year. ®

Sign up to our NewsletterGet IT in your inbox daily

11 Comments

More from The Register

Cisco and AWS hop into bed for steamy hybrid Kubernetes action

Mixing up on-premises and cloudy containers

Spooked Cisco chief phoned AWS, asked: You're not making a switch, are you?

Switchzilla's share price dipped following rumor of direct rivalry

The latest convergence whodunnit: It was Hitachi Vantara and Cisco, in the channel, flogging big iron

Networking, storage collab hitting a data centre near you

AWS will make switches to go after Cisco – report

Updated Amazon increasingly accepts hybrid cloud is a thing, but would it really get into support?

How back of server pack rat Inspur has sailed into the top-three slot

Analysis Chinese firm blows past costly Cisco and Lenovo

Kick-Kaas: NetApp gobbles cloudy Kubernetes upstart StackPointCloud

CEO to stay on as NetApp Kubernetes takes to the skies

NetApp puts MAX Data, er, ONTAP and StorageGRID gets flash acceleration

Offers a no-details flash performance guarantee

Cisco and Pure shove mini AI in FlashStack converged systems

Entry-level AIRI equivalent

Automated Weather Source didn't see this cloud coming: Amazon snatches up AWS.com

Uh, we'll be having that domain

NetApp: Public cloud, public cloud we really love the... you're going to keep some things on-premises, right? Right?

Flashes AWS and Azure cloudification, ONTAP as cloud abstraction layer