GigaIO at SC21: The International HPC Conference

Composable Disaggregated Infrastructure (CDI) is the next generation of rack architecture that makes your data center infrastructure as agile and flexible as the cloud, while reducing your costs. You gain access to more processing power and storage when you need it, and you can easily share those resources, thereby increasing their utilization. Many of these resources (like GPUs) are typically utilized only 15% of the time industry-wide. CDI will make deploying, servicing and maintaining your systems quicker and easier, and reduce administrative costs. CDI enables you to easily have different obsolescence and upgrade schedules for particular parts of your systems, eliminating “forklift upgrades”. Maximize the life of your servers while also easily integrating new accelerator technology to optimize your AI and computationally-heavy workloads. CDI enables you to do more with less.

There is not. You get the industry’s lowest latency network throughout the rack, because we never translate to and from InfiniBand or Ethernet within the rack. Whether we are connecting resources to servers, or servers to other servers, we do it all over the same PCIe network. Other vendors can achieve one of these two; only GigaIO can combine both – and you get all the benefits: less cost, less complexity, more performance.

Any workload that needs more than one server. If you can accomplish everything you need in one server, you already use native PCIe for everything. When you need more than one, or want access to more resources (storage, GPUs, FPGAs, etc.) than will fit in one server, FabreX is the perfect answer for you. Gartner adds: “where infrastructure must be resized frequently, or where composability increases the utilization of high-cost components”.

Typical workloads center around the use of GPU and FPGA accelerators, including AI/ML/DL, visualization, High Performance Computing and Data Analytics. For more use cases visit our Solutions page.

 Simply speaking an individual server has limits in terms of amount of storage, and in particular, the amount of accelerators that they can host and use (“enumerate”). NVMe-oF and GDR running natively on a PCIe network greatly increase the resources you can deploy. With NVMe-oF, you can create storage of any size, and with GDR, you can access much larger pools of GPUs or other accelerators. And with these solutions running across FabreX natively, you get the best of both worlds – full PCIe performance, AND as many resources as you need. With NVMe-oF over FabreX, you will get the performance of direct attached storage (DAS) with the flexibility of shared storage (NAS).

From a resource perspective, any PCIe-compliant device will work with FabreX because we are 100% standards-compliant. Storage products, GPUs, FPGAs, network adapters, or other add-in cards will simply work. Hundreds of companies and thousands of products support PCIe and they all work with FabreX. The vast majority of servers will also work with FabreX, as long as you have at least one x16 PCIe slot HHHL available to attach to the network. Composing resources to these servers is typically automatic and works with all operating systems, including Linux and Windows. There are however differences between servers as to the amount of resources they can support for composition. Finally, our server to server solutions require drivers to run on the server, and today these run under Linux.

No, FabreX is truly plug and play. To your application the rack-scale servers created through FabreX will appear exactly as if they were a traditional physical server, including applications that use industry standard MPI libraries. For GPU or FPGA workflows, AI frameworks are unchanged. On the other hand, you may be able to utilize more resources with FabreX than with your current infrastructure today, and you may want to make changes to take advantage of this new capability.

We provide you with freedom of choice and do not lock you into yet another proprietary software package. We start from a basic principle that FabreX is an open platform and based on industry standards – so you get all the benefits of no vendor lock-in. We strive to work with the software and environment you have already, whether that is Slurm-based, or OpenStack, or containers or virtual machines (VMs), or if you prefer to write your own scripts and control the hardware yourself. In all these cases we have great solutions either available from us or from our Ready-to-Run partners. Please visit our software section located under the Products tab for up-to-date information.

While FabreX can span distances up to 100 meters using fiber, it is uniquely capable of transforming a rack full of servers into a true rack-scale single computational element, where the rack becomes the computing element, instead of the server. To enjoy the latency, performance and composability advantages beyond a single rack, it is very easy to use one of the legacy networks, Ethernet or InfiniBand, to connect FabreX-enabled racks together. Legacy network adapter cards work seamlessly with FabreX, and it is a simple matter to configure a server to act as a gateway. We offer a pre-configured and tested Gateway product for this purpose – please see the Products tab.

GigaIO is the only CDI vendor that can transform the basic computing unit in your data center from a server, to an entire rack, using only PCIe for the absolute lowest latency and highest bandwidth throughout the rack. We want you to get the absolute most out of the investment in your IT infrastructure, and FabreX offers a unique and highly disruptive technology to transform a rack of servers into a true rack-scale system, without getting locked in with proprietary hardware or software. Making this happen requires BOTH the ability to disaggregate and compose resources to servers, AND the ability to run internode communications on the same network – we call this Cloud-Class™ composition. Composing resources with FabreX dramatically lowers your OpEx and CapEx costs, and improves your serviceability and upgradeability. And running internode communications, like NVMe-oF, GDR, MPI, Libfabric, TCP, etc., ensures easy integration and delivers scale. Vendors offering one or the other deliver less options, lower performance, more administrative hassles and more cost. Cloud-Class composition with FabreX gives you all the flexibility and agility associated with the public cloud in your own private cloud or on-prem facilities. Why spend more and get less?