Fungible: Data Processing Unit Democratizes Data Center Cloudification

World’s only workload-centric NVMe over TCP storage designed as a scalable, resilient, and programmable solution.




Modern data centers use scale-out architecture to leverage volume economics and simplify deployment. While there are many advantages to this approach, there are also pervasive inefficiencies in compute, storage and networking, limiting its broad applicability in today’s demanding environments.


The Fungible DPU was invented to address the largest inefficiencies in scale-out data centers: the inefficient execution of data-centric computations within server nodes and the inefficient interchange of data among server nodes, while strengthening reliability and security.

Leveraging this novel invention, Fungible enables a high-performance, massively-scalable hyperdisaggregated infrastructure – one that can truly rise to the call of the data-centric era.


The 41st IT Press Tour had the opportunity to meet with Eric Hayes, CEO, Brian McCloskey, Chief Revenue Officer, Reese Lloyd, Dir Product Management, and Jennifer Bell, VP, Marketing at Fungible.


Company Overview


According to Eric, Fungible's mission is to cloudify all of the world’s data centers by using the Fungible DPU to connect CPUs, GPUs, and all-flash arrays via Ethernet.

Fungible created the DPU. The company was founded in 2015 to improve the agility, security, performance, reliability, and economics of all cloud data centers.


The technology driving this is the Fungible DPU, a new class of microprocessor that powers software, systems, and solutions.


The Fungible team has developed intellectual property which includes silicon, software, systems, and solutions with more than 60 patents.


Problem/Solution


Big-fast data requires scale-out infrastructure. Fungible DPU enables data-centric cloudification. Data-centric transformations that are more than 20X faster than CPUs or GPUs. Efficient pooling of resources at scale. Robust security for applications and data. Endpoints require unique data transformations.


CPUs have evolved for 50 years and are still a $20B TAM. GPUs evolved for 30 years and are a $10B TAM. While flash memory with disruptive speed/density is a $30B TAM.


Fungible provides a complete solution from silicon to bare metal virtual data center composition software. This includes the Fungible DPU silicon, the FunOS operating system, infrastructure software embedded in the control plane with host drivers and agents, systems, and solutions for scale-out card applications and scale-out storage applications. The entire composable data center enables users to cloudify their network


Why Composable Infrastructure?


Composable promises efficient use of physical resources and enables rapid configuration of customized bare metal server and storage configurations. – Gartner Innovation Insight: Understand the Hype, Hope, and Reality of Composable Infrastrcuture, January 5, 2022
It has the potential to completely overhaul a decades-old monolithic systems architecture, which cannot scale in the digital era that is upon us. – IDC Worldwide Composable Infrastructure Forecast

Worldwide converged systems revenue from IDC shows integrated infrastructure (systems and platforms declining) while hyperconverged infrastructure grows as servers continue to be bought. Composable infrastructure is growing and has the greatest long-term future from Funglible's perspective.


Infrastructure Challenges


Diversification of workloads. The exponential growth of data. Organizations need faster response times. Legacy technologies are plateauing. All companies are looking for ways to make money off data more quickly. Process data in a simple and timely manner.

Lost time is never found. – Benjamin Franklin.

Enable faster data-driven business outcomes with scalable and agile infrastructure to process data more quickly.


The Elastic Block Storage Challenge


The cloud ideal is to have workload tuned storage, manual or automated control, single, elastic system, coordinated growth.


The non-Fungible reality is there are silos in which application-specific stacks have to be managed and controlled individually, control plane sprawl, islands, and disjointed growth.


Fungible Benefits

  • Elastic scalability whereby scale-out architecture delivers linear scalability in performance and capacity. The bigger the cluster, the better the performance and the better the efficiency.

  • Workload-centric data services where the user can define performance, capacity, and data services per volume with the benefits of centralized storage. There is flexible durability using erasure coding or raw volumes, as well as real-time compression, and encryption.

  • Consistent performance is ensured by QoS across mixed workloads. Leverage NVMe/TCP with bypass protocol conversions. Performance increases linearly as the system scales.

  • Resource efficiency with greater resource utilization per dollar spent. Optimized allocation of resources per volume for greatest efficiency. Reduces hardware required to run applications by moving data services.

  • Cloud-like agility by leveraging the API and plug-ins to simplify provisioning and allocations. Eliminates the need for administrators to best fit multiple workload demands. Flexible deployment depending on workload demands.

Fungible Storage Cluster


High performance, scale-out, and fully featured NVMe/TCP block storage. NVMe/TCP standards enable next-gen storage with 2.5 to 3.5X more IOPS, reduced latency by up to 75%, and reduced CPU load per I/O in half over iSCSI according to Dell.


Fungible DPU


A new class of microprocessor purpose-built for the data-centric era - the third socket in data centers that enables 10X more efficient execution of data-centric tasks and implements an endpoint for a more efficient data center network. The result is stateful, multiplexed processing of high-bandwidth data streams with many multi-threaded cores, hardware accelerators for crypto, compression, erasure coding, est., and native TCP offload.


Every workload is consolidated – bare metal (Oracle, Cassandra, IBM Spectrum Scale), virtualization (VMware, KVM), containers (Kubernetes, OpenShift), and OpenStack can all run side-by-side with storage tuned specifically for it directly controlled by APIs.


Multi-tenant customers are perfect prospects. Because the logic in the DPU can provide a separate encryption key for each volume on the stack. Large MSPs and telcos are especially interested in the specific workload attribute this offers.


Manageability


Interfaces in OpenStack cinder, CSI plug-in, REST intent API, and Composer GUI. The platform is self-healing, can be managed autonomously with alerts and events, and robust telemetry. All of this removes the complexity of managing storage resources. Scalability is simplified with the transparent increase of capacity and performance as needed. Users experience cloud-like flexibility in provisioning, consuming, and reallocation of resources.


Use Cases


Workload-centric data services include enterprise apps, AI/ML, large-scale databases. Allocate storage and data services based on workload requirement per volume. Allow FSC to automatically allocate resources for optimal utilization, applying optimization techniques to gain efficiency. Storage becomes a consumption-based service inside the data center.


Key Takeaways

  • Cloud-like agility to the data center. Fungible Storage Cluster allows each volume to be composed of storage and data services required by the workload.

  • Greater efficiency with 60% reduction in power draw, effective durability and security without CPU overhead, and high performance for every dollar spent.

  • Realizing the benefits of economies of scale without sacrificing economies of scope.

  • The elimination of the complexity of managing storage in a dynamic environment with Fungible Composer software that uses Intent API to determine the most optimal ways of provisioning resources.

  • The use of industry-standard plug-ins and APIs for tighter application/storage integration.