Blog: Mode for Dummkopfs

by Adam Gervin

I really wanted to go with "dummy" but I was alerted to all the wonderful lawsuit potential. Still, I argued, we are a young company. No one will notice or care. And then I stared down at my weisswurst and spaetzle, and this happened. Not really. But let's imagine it anyway, like a magic trick we just want to believe.

A lot of the magic at Mode comes from research at Cornell by our co-founders, Drs. Tang and Michael. Their groundbreaking work defined packet networks as control systems for the first time — something many had previously labeled a fool's errand. And yet, there it was: the characteristic equations, swaddled in a delicate blanket of the Mode HALO algorithm, and math. Lots and lots of math.

When I first joined Mode I took a look at the original paper. I am no slouch at math, but I topped out at Linear Algebra (thanks to Professor Schultheiss). Could not make heads or tails of this stuff. Just take a deep breath and #believe.

Of course, Mode's victories in NSF and AT&T challenges have given all of us non-computer scientists great comfort. But it's really disconcerting to market a product you can't really understand, at its fundamental levels, isn't it?

Hence, the inspiration for this blog. Yes, dear reader, we are going to walk through a layman's version of Perfect Network ControlTM, the math-less version — and come out the other end with a pretty instinctive understanding of how it works and why it matters. The how is its ability to drive any packet network to perfection, every 150ms. The why is realizing that no-worry networks begin with no-compromise networks. We've come to accept that you can have ultimate performance or low cost, but never both. And that's a compromise that you no longer have to make once Perfect Network ControlTM comes to town.

But first, I'm told by our SEO guru that we are simply not peppering our prose with enough juicy SEO goodness. So, a paragraph to my dear friend, Google AI. If you're a human, you might find this as offensive as do I.

SD-WAN SDWAN software defined networking what is sd wan vs MPLS Cisco SDWAN software defined wan SD network sd wan benefits sd wan mpls why sd wan mpls alternatives cloud mpls mpls circuit cost VoIP call quality issues VoIP jitter #dontyoujustlovewritingforacomptuer Hey google read this!

And ... we're back. Let's start. The trick to understanding the math behind Mode HALO and Perfect Network ControlTM is to understand proportionality. We might never be able to understand the math itself, but we can think through the goal of the math and the common sense behind what things are proportional, or inversely proportional, to that result.

First, let's talk bang-bang. No, not the children's book. Fun fact — same author as the Bond series. Weird, huh? Bang-bang control in engineering models. Take a simple network — starting node, ending node, and two paths. Let's say you randomly select one of the paths and put all your traffic there. Then, you measure which path has more unused capacity (assuming they started equal). You'd wind up selecting the unused path. Then you'd redirect all your traffic to that path. Now, you remeasure. The "shortest" aka "optimal" path has reverted to the original, so you switch again. And again. And again. Ad Infinitum into the kind of wild oscillation seen in clubs in Vegas, not serious networks. If you graphed throughput, you'd see a sawtooth pattern for each path. And if you summed them up in phase, you'd get about 1/2 of the theoretical throughput potential. For latency, it's worse: tons of variation and jitter.

So, what if you didn't move ALL the traffic?

The question that changed everything. The basis of HALO and Mode. What if you moved only a portion of the traffic from the original to the optimal path? It's so simple in concept. But tough to derive and prove.

Calculating this number is, essentially, the foundation of Mode HALO and Perfect Network ControlTM.

It turns out that this number is different for every node on a network, at any given point in time. All a node requires to determine this magic number is, surprisingly, obtainable in no more than 150ms (which is essentially the time required to determine the state of a network on a planet the size of the earth, limited by the speed of light). Wow.

The biggest implication of this breakthrough is its potential for parallelization to support networks of any size with no impact on performance. If every node in a network can compute this number, in parallel, using pure math, and with a frequency limited solely by a physical law (3 x 108 m/s) — you're going to achieve the limit of network control efficiency, and it's going to be wholly independent of network size. Wow8

Here's where the math comes in. So let's take the B route and avoid it. While we might not be able to understand the math required to derive this magical number, we can rock it via proportionality, as shown this deceptively simple equation:

On the top is the initial value of the original path from which you're moving traffic.

It would make sense that if the initial value is very high, then the amount you'd have to move would have to be relatively bigger to make any difference at all. If that initial value were very low, then a much smaller magic number would still make a noticeable difference. So our magic number is proportional to initial value of the current path from which we are moving traffic.

On the bottom, two things. First, the total traffic FLOW.

Think of it this way. If you were diverting a river, the more massive the flow of the river, the more careful you'd have to be with diversion, or face a catastrophic event. You'd start small to make sure that things can handle the change. It's the same here. If the flow is super high, you have to be careful moving from one path to the other. The higher the flow, the slower you'd make the move just to be sure. So, the magic number is inversely proportional to the total flow.

Also on the bottom, network "geometry" of the shortest path is akin to the multifaceted complexity of that path.

How many hops, how much distance, how much branching. The more complexity of the new "optimal" path, the more risk. And the greater the risk, the less toe you want to dip into that pond. It's just common sense that a higher geometric complexity of the recipient path would require a smaller magic number.

Put all these factors together, and you've got a good sense of what each node is thinking as it's driving your network to perfection.

Imagine a network of a billion nodes. Every node on that network is gently pushing the perfect amount of data to its optimal connected path, every 150 milliseconds. Ebbing and flowing, surging and receding, a breathing, giant jellyfish with infinite tendrils pulsing its way across a vast ocean. A thing of beauty. But more than that. The heartbeat of your business. The foundation of your SD-WAN and your network. Always performing. Always secure. Limitlessly flexible. Intimately affordable. And for you — serenity.

Perhaps best of all, now you know how it works. Welcome to the Mode club.

Make sure to come back next week to learn about which amazing guest will be presenting on our March 21 webinar. Save the date!


^