Deep Dive Into the Mode Autonomous Private Network


Today we are going to do a deep dive into Mode’s autonomous private network, following on from our post last month about their exiting from stealth and announcing Mode Core.

Co-founders Dr. Nithin Michael (now CTO) and Dr. Kevin Tang have been working on the private network for the last three years. The pair met at Cornell University’s Electrical and Computing Engineering School where Dr. Tang teaches and Dr. Michael was studying for his PhD. Michael’s thesis work focused on solving a long-standing problem in networking, using breakthrough control schemes to enable networks to be self-managing. The dissertation was the foundation for Mode Router and Mode HALO. Working together, Michael and Tang uncovered the key to the theoretical limit of network performance using characteristic equations to define modern packet-switched networks, followed by implementing a mathematically optimal system they called Mode HALO.

The startup is now headquartered in San Francisco, and the pair has been joined there by CEO Paul Dawes, who was previously CEO of icontrol, a provider of SaaS smart home products purchased by Comcast in 2017.

Mode received $24M in funding led by New Enterprise Associates, the National Science Foundation and Google Ventures to help them build Mode Core, the world’s first wholly autonomous global virtual network, which offers SD-WAN solutions at a price point the company say is dramatically lower than the cost of running software-defined wide area networks (SD-WAN) using traditional methods typically offered by telcos.

“Traditional hardware-defined private networking solutions like MPLS guarantee reliability, but are inelastic, hard to manage and costly. Mode Core was built to enhance SD-WAN, leveraging our breakthrough in routing efficiency to deliver the performance and reliability of networks like MPLS, but with the flexibility, elasticity and affordability of a cloud service,” Dawes explained in a press statement.

The Networking Problem to be Solved and the Math that Solved It

Most networks today still operate on frameworks designed over three decades ago. Technologies such as MPLS have long been used to connect business locations with their in-house applications and data. The technology was reliable and secure, but expensive and inflexible. Then cloud came along and offers more flexibility and lower price points, but the trade-off is the risk of unpredictable connectivity. SD-WAN allows businesses to manage these networks and assign different applications to different networks depending on business needs; however, typically when using SD-WAN, a choice must be made between reliability and affordability. Optimization techniques such as WANOP and TCP tuning are not always sufficient.

Mode’s own version of SD-CORE offers Quality of Service, has MPLS-like availability and comes with SLA guarantees, and is offered at an affordable price. It works alongside MPLS; alternatively, Mode Core can support the secure, phased transition from MPLS to a QoS network with an entire SD-WAN implementation that offers “an affordable, 99.99% reliable, QoS private network with SLA guarantees”.

When the Cornell researchers were defining the networking problem they focused on routing in the U.S. model. They noted that it was a form of control scheme i.e. a way to control flow through the network. Looking at control theory from an engineer’s perspective, it became clear to them both that there were insufficient feedback controls in legacy networks. The duo began to ask if they could characterize network traffic in mathematical terms and design a system that could be described mathematically in order to deploy the tools from mathematical expression that characterize behavior (previously this was thought impossible). For the first time, the pair derived the dynamical system equations that define packet-switched networks and engineered a distributed, optimal feedback control system they call Mode HALO.

Now Mode is partnering with Microsoft, Ericsson and more than 100 global service providers to apply their routing breakthrough to the huge, secure, underlay networks of their partners. These partnerships allow Mode to customize the routing infrastructure and the connections in between and have control over the underlay. The agents can make local distributed decisions because the entire system is tracking the global optimization points. The resulting global autonomous network with multiples of throughput performance leads to “the near elimination of latency variance”, making Mode CORE, “the world’s highest-performing SD-CORE”.

“Michael came up with the first math-based description of how a packet-switched network works,” says Dawes.

This allowed him to build a software-defined, automated way to route traffic on each node on the network that doesn’t need intervention to tell it how to route packets. Once Michael had that solved, it eliminated the need for more complex and expensive solutions.

In an interview with us this spring, Nithin Michael elaborated, “When you do math, it doesn’t tend to translate into the real world. But what was exciting was the protocols we had come up with were able to design a distributed feedback control system for all the network layers, as opposed to just the traditional routing network that was typically in use.”

The math discovery they made applies to layers 2 and 3 of the OSI model and cuts through previous network limits, allowing Mode Core to achieve levels of efficiency and performance previously thought impossible.

The History of Mode’s Breakthrough

In 2014, the foundation that had funded Michael’s PhD decided to compare the techniques he and Tang had developed at Cornell against what Google was using in its networks.

The experiment was carried out on the NSF network testbed, GENI. The researchers deployed a network across the U.S., using the GENI PoP to create virtual routers. They then used a randomly generated, high-demand traffic matrix to set the communication rates between PoPs, while the routers at each PoP ran HALO. The network path diversity that resulted meant that traffic at each PoP had to be apportioned in a non-trivial manner for optimal bandwidth use, minimizing overall network delay in the face of dynamic traffic spikes.

The study found that between New York and California, the Cornell self-managing networks could support about three times more traffic than Google could at the lowest possible delay, largely because of congestion issues and latency. In comparison to heuristic protocols, HALO was able to rapidly adapt to dynamic traffic changes without prior knowledge. This dynamism is the key to the optimized performance, flexibility and reliability of Mode HALO.

A subsequent win at the AT&T SDN Network Design Challenge allowed the researchers to further validate the operational advantage the NSF experiments had revealed.

The researchers presented the study’s findings at Nanog, and found that many VCs in Silicon Valley were excited that the long-standing problem could be solved in this way and further convinced by how feasibility was demonstrated by the Cornell computer scientists. Investors saw tremendous business potential in their approach.

In the summer of 2015, Michael and Tang formed an NDA. In the spring of 2016, they raised $8.3M in Series A funding with NEA as the lead investor and a grant from the National Science Foundation. More recently, they achieved a $16M Series B round, bringing their total fundraising to $24.3M to date. The challenge for what was then called Waltz Networks was how to bring this to market in way that could rival an incumbent like Google? Michael says the answer came to them as they started to talk to potential customers about how best to provide their services.

Mode CORE: Key Innovations and Breakthroughs

  • A new pure math backbone for a post-HTTP world
  • Near elimination of latency variance and jitter (99.5% of problems happen in the first and middle miles of the core)
  • High performance for UC, SaaS, remote access and site-to-site
  • A breakthrough math discovery powered the creation of Mode CORE
  • The breakthrough replaces the current routing state-of-the-art layers 2 and 3 of the OSI model
  • Mode can then define and deliver the theoretical limit of network routing
  • Mode’s pure math routing algorithm is applied to a mix of Microsoft and Ericsson’s global underlay networks
  • Worldwide core network brings ultimate UCaaS performance globally for hundredths of a cent per connection minute
  • One-step SD-WAN device configuration offers the ability to send your traffic to the closest Mode Core PoP globally
  • Each virtual router in the Mode CORE network consumes two CPU cores – one for the containerized control plane and one for the Open Virtual Switch (OVS)-based data plane
  • A single virtual router using a small amount of the available server CPU capacity delivers over 6 Gbps of throughput, meaning that in a multi-core server, Mode is able to achieve line rate throughout from the NICs (typically 40 Gbps)
  • Traffic is routed securely across Mode’s optimized private core to either unified communications solutions, cloud services, or remote users on VPNs
  • Every virtual router in the Mode CORE can support up to 100K flows/sec while storing 100K flow rules in the flow table – up to par for most enterprise VPNs
  • Businesses can spin up a private network in less than 60 seconds, bringing with it private network security
  • The number of private networks is unlimited
  • Separate networks can be used for separate applications
  • QoS bandwidth is automatically adjusted in real-time to satisfy dynamic demand across all the Core networks you have in play
  • Private networks are transparent end-to-end, allowing you to manage your WAN in the same way as your LAN using the Mode web portal tool

Mode CORE: Use Cases

Mode CORE is particularly beneficial for real-time applications that need ultra-low latency (ULL) performance, such as:

  • Interactive livestreaming applications
  • Real-time applications that require reliability
  • Multiplayer gaming
  • Real-time machine learning
  • Remote command and control
  • Feedback-loop AI

It is also suitable for:

  • Site-to-site connectivity
  • Remote access
  • All cloud applications


Mode provides a dashboard in real-time to show how traffic is dynamically adjusting.

Scroll to Top