データセンターへの代替可能な目標到達プロセスの構成可能なインフラストラクチャ

Fungibleは、本日Fungible Data Centerプラットフォームを発表し、企業およびサービスプロバイダーのデータセンターに構成可能なインフラストラクチャを提供したいと考えています。Fungible wants to bring composable infrastructure to enterprise and service provider data centers with the launch today of its Fungible Data Center platform. “Our goal is to bring the performance, efficiency, and simplicity that hyperscalers have historically enjoyed to data centers of all scales, anywhere from half a rack to hundreds of racks,” Srinidhi Varadarajan, VP of solutions at Fungible, said. With the launch of the Fungible Data Center, the company’s goal is to breakdown traditional data center conventions where a workload runs on a specific server SKU to one where CPU, GPU, or storage resources can be pooled across multiple racks and allocated to workloads on the fly. “One of the things we noticed in modern data centers is there are a very large number of different server types, or servers SKUs,” Varadarajan said, adding that these servers are often built to run specific workloads. The problem with this is when a server is under load, it’s not possible to reallocate resources from underutilized servers elsewhere in the data center, resulting in poor overall utilization of the available resources, he explained. The Fungible Data Center addresses this challenge by enabling users to pool idle resources from any server on the network to run workloads. “We now have a mechanism to deploy and manage a turnkey multi-tenant data center with single pane of glass management and be able to deploy workloads … without any application changes,” Varadarajan said. Rather than selling individual server components, Fungible is offering server racks pre-populated with the customer’s choice of CPU, GPU, and storage resources. According to Varadarajan, a typical rack might be equipped with a Juniper top-of-rack switch, any number of Fungible’s FS1600 Storage Clusters and Composer appliances, as well as standard CPU-or GPU-centric servers. These resources are then stitched together using the company’s data processing units (DPUs). Fungible’s FS1600, which was announced last fall, features dual F1 DPUs each capable of 800 Gb/s of throughput, while each server is equipped with the company’s Fungible Data Services PCIe card running an S1 DPU capable of 200 Gb/s of throughput. The DPUs then interface with the rest of the rack over 50 Gb/s, 100 Gb/s, or 200 Gb/s Ethernet running over standard transmission control protocol (TCP). Fungible’s DPUs behave like smartNICs and are tasked with offloading high input/output (I/O) workloads and services from the CPU. The DPU forms a network fabric that enables workloads to address idle CPU and memory resources on one rack, GPU resources on another, and storage resources on yet another. If “a workload comes along and says, I want the server to look like this, you can borrow, from each of the pools, the corresponding parts that are needed to make that server, and build it dynamically, and do this in minutes,” Varadarajan said. “You can now move resources from one workload to another workload on the fly, eliminating bottlenecks.” Varadarajan notes that because the company’s DPUs operate over TCP and behave as if they were standard Ethernet NICs, customers aren’t locked into the platform. Workloads are managed via the Fungible Composer software platform, which handles the composition, provisioning, and management of resources across the data center. For example, if a customer needs to spin up a GPU-intensive workload, the Fungible composer pairs the workload with the appropriate compute server and automatically provisions storage and GPU resources from anywhere in the data center. And according to Varadarajan, this isn’t a rigid configuration. Those resources can be dynamically re-allocated to other workloads as demand changes throughout the day. If more resources are required, the composer will add additional storage, GPUs, or CPUs to the pool, while if there is a lull in demand, those resources can be given to another or higher priority workload. “The beauty of it is you don’t have to buy and keep hardware idle,” he said. “You can simply use servers that are idle in one place as your virtual queue, and this can be done in minutes.” To simplify this process, Fungible has created a hierarchical marketplace, akin to what you’d see from one of the public cloud providers, which defines the minimum resources to assign a specific workload. Varadarajan adds that the Composer not only integrates with platforms like Ansible Playbooks and HashiCorp’s Terraform, but is also multi-tenant, allowing managed service providers to assign resource pools to specific customers. Fungible’s Data Center offering is aimed at service providers that want to compete with hyperscalers, enterprises looking to simplify data center management, and university research labs with limited resources. The platform can be had in three sizes ranging from a half-rack configuration on the low end to multiple full-height racks on the high end. CPU and storage pools are available now, with GPU pools on the company’s roadmap for the near future. In addition to the upfront hardware cost, pricing is based on a subscription model that takes into account the number of servers and storage clusters managed by the Fungible Composer.