Co-founded in 2015 by Pradeep Sindhu, one of the co-founders of Juniper Networks, and Bertrand Serlet, former head of software engineering at Apple, Fungible continues to develop its block mode platform offering companies to “cloudify their primary storage”. As we already explained to you last November, the start-up uses its DPU F1 cards (data processing unit) to speed up the processing of NVMe/TCP requests in a storage array. Recall that a DPU (Data Processing Unit) offloads the host system processor from managing I/O traffic to accelerate the performance of a storage platform. That of Fungible claims capacities of 13 million IOPS – against 10 previously – with a latency of 100 microseconds and a throughput of 75 GB/s per node. Storage capacity starts at 70TB – with a minimum of three nodes – and goes up to several petabytes. While GPUs are very good at accelerating the execution of massively parallel applications, CPUs still stand out for the execution of traditional applications with a single thread and DPUs are above the rest for performing encryption functions, as well as data packet processing for network and storage.
In the FS1600 chassis, 2U rack-type x86 servers, the F1 FC200 or FC50 DPUs – based on a MIPS64 chip (with 64 cores) and a PCI 4.0 interface – drive NVMe SSD drives (2×12) with 2 links respectively 100 GbE or 25 GbE per card. Exploiting the resources of the SoCs fitted to the DPU cards, the in-house operating system, FunOS, provides encryption, compression, thin provisioning, snapshots and QoS functions in erasure coding mode, and orchestrates the entire storage stack with the networking, security and virtualization (SR-IOV). Fungible’s TrueFabric software, an IP fabric IP over Ethernet protocol, is embedded on-chip and provides, according to the vendor, cross-bandwidth with low latency, end-to-end QoS and congestion-free connectivity. This enables disaggregation and pooling of all storage cluster resources. TrueFabric is interoperable with TCP/IP over Ethernet, ensuring that a data center’s spine-leaf network based on standard Ethernet switches is immediately ready for use.
Two DPUs in an FG1600 chassis supports up to 24 NVMe SSDs with performance of 4 million IOPS( 2 per DPU). (Fungible Credit)
With its disaggregated and scalable solution, Fungible explains that it solves two of the main challenges in data centers: to execute calculations more efficiently between server nodes in order to exchange data more quickly within a storage cluster. To drive its scale-out NVMe over Fabrics (NVME-oF) storage nodes, the Fungible Storage Cluster (FSC) arrives in version 4.0 and ensures separation of storage control planes (control plane) and data planes ( dataplane). The platform is designed to enable enterprises and cloud service providers to achieve hyperscale performance in their data centers. FSCs are the foundation of the vendor’s Fungible Data Center (FDC), a turnkey, on-premises solution comprised of pre-configured racks with disaggregated compute and storage servers powered by DPU boards and managed by Data Center Composer (DCC) software ).