Smile – Fungible’s FunOS 4.0 brings DPU delight to entry-level systems

Adding server-offload cards to smaller numbers of servers has become possible with Fungible’s latest SW release.

Fungible is a Data Processing Unit (DPU) startup with two proprietary S1 and F1 chips that run its storage initiator cards and its scale-out  FS1600 NVMe array. The S1 and F1 DPUs run FunOS software on their data plane and it  features networking, storage, security, virtualisation, and analytics functionality.

The control plane runs Linux and contains agents that allow a cluster of F1 and S1 DPUs to be managed, controlled, and monitored by a set of REST APIs.

Fungible this week announced v4.0 FunOS to Blocks and Files, which adds support for smaller entry-level systems with reduced erasure coding and replica schemes. Chief Revenue Officer Brian McCloskey said: “One of the great things about the release of the 4.0 software stack is it provides erasure coding for smaller configurations. The sweet spot starter configuration for us prior to 4.0 was a six node cluster to enable 4 + 2 erasure coding. With the release of 4.0 it drives data efficiencies up and improves price per terabyte at smaller scale.” 

Erasure coding (EC) is an alternative to RAID and replication schemes, it has less overhead needed for the protection parity data with RAID and the copies needed with replicas. With erasure coding a data item is split into sub-sections with mathematical descriptive information (parity block) and each stored in a separate drive – or FS1600 node in Fungible’s scheme. If a storage node or drive fails or the data becomes corrupted, then the whole data item can be reconstructed from the other fragments with their parity blocks.

A 2+1 EC scheme divides each unit of data into two blocks with an added parity block, meaning 3 Fungible nodes are needed. A 4+2 scheme involves four data blocks and 2 parity blocks (6 nodes), while an 8+2 setup has 8 data blocks and 2 parity blocks (10 nodes).

The Fungible slide above illustrates that Fungible’s erasure coding scheme uses less node capacity overhead, than double or triple replicated alternatives protecting against single or double drive failure. The v4.0 release means Fungible kit can be used with 3-node clusters instead of needing 6 nodes.

McCloskey said: “The feature that we’re adding as part of FSC 4.0  is erasure coding 2 plus 1. We wanted to [have an] erasure coding level that lets a customer start small with our system, and then evolve and grow over time.”


Fungible also offers double replicated (RF2) volumes for extra data durability. The V4.0 SW brings in single replicated volumes which can be compressed. Fungible’s director for product management, Reese Loyd, said: “We’re able to layer compression and other data services on top of that. So the big one that we’re announcing as part of 4.0 is compression. You’ll be able to leverage single replica volumes that are both thin provisioned [and] compressed.”

Fungibles says it captures a lot of the raw performance of the NVMe drives in its FS1600 storage cluster, while also offering the compression data service. This can be used to extend application durability and to protect containerised workloads where speed is paramount.

Replication or erasure coding

Coldago analyst Philippe Nicolas told us: “Replication consumes lots of storage with x2 x3 the original size but it has very easy semantics. It is very good for small content, but when blocks become large it can create a penalty.”

“EC introduces striping plus protection (parities) and can accelerate throughput as multiple nodes collaborate, it also reduces the HW overhead compared to replication.” With Fungible’s dedicated DPU it’s pretty fast and very good for large blocks.

Replication requires fewer nodes and this makes it useful for disaster recovery.

He said applications such Cassandra come with built-in protection with 3 replicas at their level and don’t rely on sophisticated storage. Fungible announced RF1 to avoid having its own 3 copies at Cassandra level and redundancy at the cluster level.

“This comes from the fact that these applications were designed on commodity HW essentially with server-based storage (internal/DAS storage) in a shared-nothing model and they required some level of redundancy at the application level.”

Nicolas added: “RF1 … means that Fungible targets the kind of highly scalable databases used by hyperscalers and large corporations: Cassandra, MongoDB and they’re open source.”

“Keep in mind for such DB engines; they implement replication to avoid split brain problems and maintain data consistency.” With Fungible “they can still serve data with fewer nodes.”

Storage Initiators

Fungible announced its Storage Initiator cards in September last year and they need FunOS, but FunOS v4, with Storage Initiator support, is only being announced now.

Lloyd told us: “We announced the Storage Initiator in the fall and had functional internal code as well as using the functionality in our FDC product line. That is what allowed us to showcase the functionality at SDSC and others. [FunOS] 4.0 productises the feature fully for the Fungible Storage Cluster (FSC) and marks the general availability.”

Further: “The FSC 4.0 release represents the full product release including FunOS on the DPUs (both FAC cards and FS1600 nodes) and our composer (UI, API, telemetry, etc) software that runs on our composer nodes. This full package is FSC 4.0.”

FAC stands for Fungible Accelerator Cards and FDC is the Fungible Data Centre.


Lloyd said Fungible will “be able to layer additional data services on top of in the future.”

Fungible Storage Initiator cards.

The future may see Fungible’s storage initiator cards supporting non-Fungible storage targets. CEO Eric Hayes said: “We have no announcement on that yet. But we are very much supportive of supporting an open standard as NVMe over TCP for interoperability… Integration on the control plane, just like anything else, is where the work needs to be done. And we are entertaining this right now. We’re very supportive of these ideas.”

Another coming feature will be around composability. McCloskey said: “You’ll see an announcement from us in early April around our next step in the composable infrastructure market.”


Fungible’s boss said the company was making progress selling its products to customers. ”We have paying customers. Two of the world’s 10 biggest hyperscalars are customers. We have service provider customers. One of the world’s biggest retailers… is a customer.”

He also mentioned a very large telco, content delivery networks, and multiple US government entities as being Fungible customers. The company is hoping to drum up greater interest at large scale prospects in the aforementioned vertical markets, as well as with SaaS and managed service providers and larger enterprises.

It’s also looking for OEM opportunities.


The company is moving quickly, and it knows it has to do so. Intel is pushing its own version of the DPU as is NVIDIA with its BlueField (very) SmartNIC devices. These can run the vSphere hypervisor and thus provide features through added VMs. 

Intel and NVIDIA are established companies with strong channels and market presence. Fungible, as a startup has to punch well above its weight to make progress. It has a good base with its fast hardware and is building up its software features as rapidly as it can.