Edge & 5G
Gain crucial understandings of Edge software and 5G concepts with Intel® industry experts
92 Discussions

Building Cloud-to-Edge Programmable Networks with Intel Innovation

Ed_Doe
Employee
1 0 5,016

The network of the future won’t look like the network of the past two decades. Sure, it will have switches, routers, optics, and NICs, but the network of the future will be programmable and intelligent with deep visibility into the control and data planes. In this blog, I’ll explore some of the networking challenges facing today’s enterprises, cloud service providers, and communications service providers, then explain Intel’s vision for the network of the future and how we’re helping turn that into reality today.

 

Networking Challenges Raise Some Big Questions

Several trends are converging to make the traditional, fixed-purpose “black box” network woefully inadequate for today’s networking requirements. These trends include:

  • New architectures. Enterprises at large are migrating to cloud-native architectures with container-based processing, automation and orchestration with the expectation to reduce their capital expenses (CAPEX) and operational expenses (OPEX).
  • Data explosion drives need for more bandwidth. Experts at IDC predict that the global datasphere will grow by 181% from 2020 through 2025 (reaching 180 ZB),[1] with at least a quarter of that data being real-time data.[2] This will drive an increased need for data to be processed at the edge to meet latency requirements. Gartner anticipates more than 50% of enterprise-generated traffic is expected to be at the edge by 2025.[3]
  • Distributed, scale-out workloads create dynamic networks. The distributed, scale-out world mandates a need for a fundamental network architecture evolution so that the network can adapt to these new software applications. In particular, the “AI Spring”—where artificial intelligence (AI) and machine learning (ML) are being adopted in every industry—means that more data needs to be processed by clusters at much faster speeds and network performance must be tuned to specific, continually evolving AI and ML workloads.
  • Increased visibility needed to gain insights into new workloads and unleash increased performance. To address the rise of new workloads, operators need deep visibility to tune networks to enable workloads to reach their peak performance and ensure maximum reliability and availability.

Other network concerns include maintaining end-to-end security and how to evolve the network while minimizing CAPEX and OPEX.

A Smarter, More Performant Network Provides the Answer

As part of the data center of the future, Intel is helping build the network of the future that can solve the aforementioned challenges. The overall vision includes Intel® Intelligent Fabric (see figure below), which accelerates enterprise, multi-cloud, 5G, IoT and HPC/AI workloads. (To clarify, “fabric” refers to a leaf/spine topology with optimized connections.) I’ll talk a bit about some of the Intel Intelligent Fabric components and benefits in the next section.

Fig1.jpg

Intel Intelligent Fabric Components

Let’s take a closer look at how Intel’s portfolio of network connectivity empowers the network of the future.

  • Infrastructure processing units (IPUs) for higher security and performance. The IPU is an advanced networking device with hardened accelerators and Ethernet connectivity that accelerates and manages infrastructure functions using tightly coupled, dedicated, programmable cores. In multi-cloud environments, IPUs isolate these maintenance functions from customer code. IPUs can also reduce overhead and free up CPU performance for revenue-generating tasks. Customers can use an IPU to better utilize resources with a secure, programmable and stable solution that enables efficient utilization of data center resources.
  • Fully programmable, intelligent Ethernet switches with In-band network telemetry (INT). With the advent of Intel® Tofino™ Intelligent Fabric Processors (IFPs), fully programmable Ethernet switches can provide the intelligence, performance, visibility and control that complex networks need. Intel Tofino-based switches generate telemetry data that can feed into Intel® Deep Insight Network Analytics Software to create reports that provide greater potential for acceleration, optimization and flexibility.
  • Optical modules for higher bandwidth. Intel® Silicon Photonics combines the silicon integrated circuit and semiconductor laser. It enables data to transfer faster over longer distances compared to traditional electronics, while utilizing the efficiencies of Intel’s high-volume silicon manufacturing. By deploying this technology, you can achieve high-bandwidth connectivity at 100G, 400G, 800G[4] and beyond.
  • For CPUs and xPUs. CPUs are for general-purpose compute, while xPUs—like graphics processing units (GPUs), field-programmable gate arrays (FPGAs) and vision processing units (VPUs)—are for application-specific or workload-specific acceleration. These components complement the IPUs for overall data center modernization and acceleration.
  • Industry standards. Intel has taken a leadership role in developing industry standards designed to reduce overhead and uplift networking performance. For example, in addition to the P4 programming language, Intel Intelligent Fabric can utilize the Linux Foundation’s Software for Open Networking in the Cloud (SONiC) OS, Extended Berkeley Packet Filter (eBPF), Storage Performance Development Kit (SPDK), Data Plane Development Kit (DPDK), External Border Gateway Protocol (EBGP) and the new Infrastructure Programmer Developer Kit (IPDK).This new kit is an open-source framework of drivers and APIs for infrastructure acceleration and management that runs on a CPU, IPU, data processing unit (DPU) or switch.

Intel Intelligent Fabric Benefits

  • Ease of use. Easy to add, modify and remove workloads via network environment orchestrator.
  • Massive bandwidth. Up to 25.6 Tbps switches with Intel Tofino 3 IFPs,[5] up to 2x100 Gbps Intel® Ethernet Network Adapters[6] and up to 400 Gbps Intel Silicon Photonics optics.[7]
  • AI/ML workload acceleration. Up to 2.27x performance improvement in ML training using SwitchML technology on a 100 Gbps Ethernet network.[8]
  • End-to-end security. Watch the Day 2 Intel Vision keynote from my colleague, Greg Lavender, SVP, CTO and GM of Intel’s Software and Advanced Technology Group, to learn more about Intel's comprehensive approach to software security, including open-source software through the Open Source Security Foundation.
  • Automation. Self-monitoring/self-analyzing/self-healing capabilities.

We envision accomplishing the above while improving density and optimizing power consumption and cost.

 

TCP-INT Demo

As one example of the benefits Intel Intelligent Fabric can provide we presented a demo, at Intel Vision,  on how embedding In-band Network Telemetry (INT) data into Transmission Control Protocol (TCP) headers can not only greatly improve visibility for congestion control, but also provide up to a 2.5x increase in scale/performance of an existing workload with the same infrastructure, by optimizing the network to adapt to the application.

tcp-int-diagram.png

 

In this scenario using TCP-INT can enable up to 2.5x more requests per connection than existing algorithms like Data Center TCP (DCTCP) that uses Explicit Congestion Notification (ECN) to estimate the degree of congestion.[9]

Ready to Start Building Cloud-to-Edge Programmable Networks?

If the network intelligence, performance, visibility and control that I’ve described here sounds like something you’d like to explore more, consider joining and contributing to some of the projects I’ve mentioned, like P4, SONiC, and IDPK. For more information about how Intel is working with the ecosystem to develop the network of the future, watch the Intel Vision video, “Cloud-to-Edge Programmable Networks.”

 


[1] Statista, “Volume of data/information created, captured, copied, and consumed worldwide from 2010 to 2025,” March 2022.

[2] IDC Report, “Revelations in the Global DataSphere,” July 2021. The creation, replication, and need of real-time data took a hit during the pandemic, but it is still projected as 25% by 2025.

[3] Gartner, “Predicts 2022: The Distributed Enterprise Drives Computing to the Edge,” October 2021. 

[4] Gazettabyte, “Intel details its 800-gigabit DR8 optical module," https://www.gazettabyte.com/home/2021/6/29/intel-details-its-800-gigabit-dr8-optical-module.html

[5] Intel® Tofino™ 3 Intelligent Fabric Processor Product Brief, http://www.intel.com/content/www/us/en/products/network-io/programmable-ethernet-switch/tofino-3-brief.html

[6] Intel® Ethernet Network Adapter E810-2CQDA2 specifications, https://ark.intel.com/content/www/us/en/ark/products/210969/intel-ethernet-network-adapter-e8102cqda2.html

[7] Intel® Silicon Photonics webpage, https://www.intel.com/content/www/us/en/products/network-io/high-performance-fabrics/silicon-photonics.html

[8] For full configuration and testing details, refer to the following paper: A. Sapio, M. Canini, C. Y. Ho, J. Nelson, P. Kalnis, C. Kim, A. Krishnamurthy, M. Moshref, D. Ports, & P. Richtarik. “Scaling Distributed Machine Learning with In-Network Aggregation.” In 18th USENIX Symposium on Networked Systems Designand Implementation 2021 (NSDI 21).

[9] Details at intel.com/PerformanceIndex , Event - Vision 2022, Session 68.

About the Author
Edward (Ed) V. Doe is vice president of the Connectivity Group and general manager of the Barefoot Division (BXD) at Intel Corporation. Doe holds a bachelor’s degree in electrical engineering from the University of Western Ontario, London, Ontario, Canada and two years of graduate studies towards a master’s degree in electrical engineering at Stanford University.