The lure of composable infrastructure is that it allows the enterprise to scale hardware footprints up and out even as it supports the added flexibility of fully software-defined architectures. But all of this adding and removing of compute/storage modules can only happen if the system is anchored by a flexible network fabric.
This represents a vastly different operational paradigm from today’s fixed network topologies.
Fabric networking is nothing new, of course, but fabric networks have typically been limited to processor interconnects and, more recently, rack-level and storage area networking. Under the composable paradigm, however, the data center itself essentially becomes a large, distributed mainframe. The basic PCIe or Infiniband interconnect can now produce fabric-style connectivity between compute and storage modules, and those fabrics should be able to integrate pretty cleanly with the connectivity solutions between individual cores within those modules.
A case in point is the new Liqid Grid that Colorado firm Liqid is planning to release in a few months as part of its Liqid Composable platform. According to eWeek, the system uses an intelligent PCIe fabric switch to provide bare-metal composable ecosystems on static, legacy infrastructure. The system boasts 24 ports of PCIe Gen 3.0 switching, configurable to x4, x8 or x16, with full duplex bandwidth of 192 GBps and latency less than 150 nanoseconds. At the same time, the company’s Liqid Command Center software provides both fabric and bare-metal management and orchestration, as well as status monitoring for clusters, machines and devices — all accessed via multiple control methods, including a GUI and RESTful APIs.
Meanwhile, One Stop Systems (OSS) is leveraging its knowledge of high-performance computing (HPC) to build composable architectures around Nvidia GPUs and other advanced processors. The company says it can connect processing and NVMe-based storage systems using a single PCIe fabric in support of artificial intelligence, deep learning and emerging peer-to-peer communications solutions. This allows organizations to upgrade server, storage and other resources on different schedules to produce a more tailored environment than an all-in-one modular solution.
Fabric-based composable infrastructure is also expected to dovetail nicely with containerized workflows, giving enterprises an unprecedented level of flexibility in crafting next-generation data services. As The Next Platform explained recently, Red Hat is looking to incorporate both Docker containers and the Kubernetes orchestration stack within its OpenShift platform, and it will then link that to the new composable networking features in OpenStack Platform 12. This will allow users to define their own network topologies rather than select from predetermined options and even integrate them with third-party solutions, such as Intel’s Rack Scale Design, through the Distributed Management Taskforce’s new Redfish API.
Traditional infrastructure companies like HPE are also shoring up the fabrics for their new composable platforms. Over the summer, Mellanox got the nod to support the first HPE Synergy Switch Module with its Spectrum Ethernet ASIC. The device provides 25, 50 and 100 Gbps connectivity, giving the enterprise not only a means to upgrade throughput across the data center, but on new edge infrastructure as well. The Spectrum offers full line rate connectivity, zero packet loss and latency on the order of 300 nanoseconds, which the companies say will support advanced analytics and virtual networking.
It is, of course, possible to build composable infrastructure around standard, non-fabric topologies, but for all practical purposes an integrated, interconnected networking environment is the way to go. Part of the appeal of a composable solution is the ability to plug in new modules so as to add resources to available pools quickly and easily. And you can’t very well do that using traditional provisioning and fixed connectivity patterns.
Going forward, data traffic will require not only high speed but broad flexibility to meet the demands of an increasingly digital-facing economy. For the moment, the only way to get there is through fabric-style networking.
Arthur Cole is a freelance journalist with more than 25 years’ experience covering enterprise IT, telecommunications and other high-tech industries.
