Let the SAN shine: Nebulon drops cloak, reveals DPU-enhanced, cloud-managed server SAN

Startup Nebulon has come out of stealth to reveal scale-out, on-premises, server SAN, block-based storage using commodity X86 servers bolstered with storage processing offload cards, along with a data management service delivered from its cloud.

It claimed its so-called cloud-defined storage (CDS) is less pricey than equivalent all-flash SAN array storage and doesn’t use up CPU capacity in its host servers, a disadvantage it claims affects both software-defined storage (SDS) and hyperconverged infrastructure (HCI) systems using commodity server chassis.

Siamak Nazari.

A prepared quote from Siamak Nazari, co-founder and CEO of Nebulon, said: “Cloud-Defined Storage delivers global insights, AI-based administration and API-driven automation making enterprise-class storage a simple attribute of the data centre fabric with self-service infrastructure provisioning and storage operations as-a-service for application owners.”

Nebulon’s storage is embodied in its Storage Processing Unit, an add-in, FH-FL PCIe card, with an 8-core, 3GHz ARM CPU plus encryption/dedupe offload engine, that is layered in front of a host server’s SAS or SATA SSDs, and connects to them via a triple SerDes connector. Nebulon co-founder and COO Craig Nunes told B&F the SPU will support “NVMe when it becomes generally available in the early Fall.”

The SPU card, which looks to upstream system software like an HBA or RAID card, presents block volumes to applications running in the servers. Up to 32 servers can be clustered in an nPod with the SPUs connected by a 10 or 25gigE network. There is a separate 1gigE port for management from the cloud.

Nebulon SPU

Data services provided by the SPU include deduplication, compression, encryption, erasure coding, snapshots and mirroring. There is no GPU on the card.

The SPU contains 32GB of NVRAM to speed writes, and reads come straight from the SSDs. NVRAM write caching means the SPU can turn random writes into sequential writes to the SSDs, thus helping to lengthen the drives’ endurance. Data is not striped across SPUs.

Initially 4TB SSDS are supported, with up to 24 supported by a single SPU, meaning 96TB, and a maximum capacity of 3,072TB across the 32 SPUs in an nPod. There is a single, all-flash storage block-access tier.

The performance on reads is slower than if NVMe SSDs were supported. Nunes told us: “At the device level, SATA latencies can be many times [that of] NVMe. However, when measured with the enterprise data services software stack, the latencies at the application level will be in the 300us to 400us range and acceptable to the cloud native, container and hypervisor use cases we are targeting.”

OEM channel

Nebulon sells its SPU and ON service through an OEM channel, with both HPE and Supermicro signed up so far, and a third OEM likely. An HPE configuration is based on the ProLiant Dl380 gen 10 server in a 2U x 24 slot chassis, while Supermicro uses its Ultra line of servers for Nebulon storage 

That means that actual server hardware configurations, including drive types and capacities come from the OEMs. So too do purchase and/or subscription arrangements.

Nebulon is pitching its product, through its OEMs, at mid-to-large enterprises needing block storage at PB and up scale and wanting to increase storage and app server efficiency and reduce acquisition and management costs.

The card can be set up using application templates to optimise it it for different workloads, such as VMare, MongoDN and Kubernetes. Nebulon storage supports any OS or hypervisor in the host server. NebOS upgrades are non-disruptive.

SPU management

The SPU runs NebOS software and is managed through a Nebulon ON SaaS service hosted in a Nebulon cloud which uses multiple CSPs and multiple regions for high-availability. It is updated through the ON service.

Nebulon says the ON service manages fleets of Nebulon systems at scale. These systems send telemetry messages to the ON Cloud; tens of thousands of storage, server and application metrics per hour. These are stored in a distributed time series database. 

ON includes an AIOps function which looks at the telemetry, analyses it in real time, and responds to adverse events in seconds by re-jigging a Nebulon system to respond to changing operational patterns. It also provides storage usage metrics over time and predictive analytics.

Nebulon ON dashboard

Customer admin staff can self-provision Nebulon storage through an ON dashboard and ON can deliver automated updates across a Nebulon fleet. 

Replication will be delivered as a future upgrade, possibly in the next software release. We expect stretch cluster support in a future release as well.

The SPU is a DPU

Nebulon said the SPU card is an example of a DPU (Data Processing unit), a dedicated storage or networking processor intended to offload storage and/or network processing from a host server’s CPU so that it can concentrate on application processing.

Examples of DPU supply and use include Diamanti, Fungible, Pensando, Nvidia (SmartNICs) and AWS for in-house use (Nitro). 

We might say Nebulon is an HCI system on DPU steroids; so many that Nebulon claims it is no longer an HCI system at all, but an ultraconverged system.

The SPU is a gateway to the storage for its host CPU. If it fails, the host server loses access to its storage. A server can have other storage installed which is not accessed through the SPU. A loss of internet connectivity will not prevent an SPU from functioning.


The Nebulon system, being a kind of super-server SAN system, will compete with Dell’s VxRail and Nutanix systems. It will also compete with disaggregated HCI systems such as Nimble’s dHCI and Datrium.

Inside HPE the Nebulon storage competes with or complements SimpliVity HCI and Nimble dHCI systems, and Primera, 3PAR and Nimble arrays. It also competes/complements other HPE storage partners providing block array services such as Datera.

Nebulon is not supported by HPE’s cloud-delivered, predictive analytics Infosight management or its GreenLake subscription service, but it is early days.

Nebulon’s Craig Nunes says more than half of HPE’s servers are sold into customers with non-HPE storage. The Nebulon storage, which should cost less than external array storage, and uses lower server SAN CPU resources, gives HPE a win-back opportunity in his view.

Regarding Pensando Nunes tells us: “Each Nebulon SPU has an 8-core 3GHz processor and 32GB of battery backed NVRAM, and runs the entire software stack you might find on a 3PAR or Pure Storage array controller.  As a compare, Pensando supports 8GB of RAM—enough for the network/security functionality but not enough to run a full storage SW stack on the card.”  

Executive Chairman David Scott says Nebulon and Pensando are complimentary: “I could easily see some use cases where a customer has both a Pensando DSC card and a Nebulon SPU in the same application server(s). :)”


Nebulon is entering a new and undeveloped market, the DPU-enhanced server SAN market with cloud-delivered, AIOps management, and its competitors, at the OEM level, are suppliers such as Pensando and the other DPU suppliers.

At the end-user level its competitors are, well, legion, and existing SDS and HCI vendors will say Nebulon is simply just another SDS or HCI vendor, one using proprietary HW and SW to give its host server chassis a performance kick. If customers accept that positioning then suppliers will compete on speeds, feeds, support – the usual stuff.

If customers see Nebulon as a new class of server SAN, then the OEM+Nebulon offer will be differentiated, although this will require a marketing and sales push.