Rethinking the Internet of Things

Front matter

Rethinking the Internet of Things

A Scalable Approach to Connecting Everything

Francis daCosta Forewor d by Alok Batra, CEO, MQldentity, Inc

Introduction

I didn’t set out to develop a new architecture for the Internet of Things (IoT). Rather, I was thinking about the implications of control and scheduling within machine social networks in the context of Metcalfe’s Law. The coming tsunami of machine-to-machine interconnections could yield tremendous flows of information - and knowledge.

Once we free the machine social network (comprised of sensors and an unimaginable number of other devices) from the drag of human interaction, there is tremendous potential for creating autonomous communities of machines that require only occasional interaction with, or reporting to, humans.

The conventional wisdom is that the expansive address space of IPv6 solves the IoT problem of myriad end devices. But the host-to-host assumptions fossilized into the IP protocol in the 1970s fundamentally limited its utility for the very edge of the IoT network. As the Internet of Things expands exponentially over the coming years, it will be expected to connect to devices that are cheaper, dumber, and more diverse. Traditional networking thinking will fail for multiple reasons.

First, although IPv6 provides an address for these devices, the largest population of these appliances, sensors, and actuators will lack the horsepower in terms of processors, memory, and bandwidth to run the bloated IP protocol stack. It simply does not make financial sense to burden a simple sensor with all of the protocol overhead needed for host-to-host communications.

Second, the conventional implementation of IP protocols implies networking knowledge on the part of device manufacturers: without centrally authorized MAC IDs and end-to-end management, IP falls flat. Many of the hundreds of thousands of manufacturers of all sizes worldwide building moisture sensors, streetlights, and toasters lack the technical expertise to implement legacy network technology in traditional ways.

Third, the data needs of the IoT are completely different from the global Internet. Most of the communications will be terse machine-to-machine interchanges that are largely asymmetrical, with much more data flowing in one direction (sensor to server, for example) than in the other. And in most cases, losing an individual message to an intermittent or noisy connection will be no big deal. Unlike the traditional Internet, which is primarily human-oriented (and thus averse to data loss), much of the Internet of Things traffic will be analyzed over time, not acted upon immediately. Most of the end devices will be essentially autonomous, operating independently whether anyone is “listening” or not.

Fourth, when there are real-time sensing and response loops needed in the Internet of Things, traditional network architectures with their round-trip control loops will be problematic. Instead, a way would be needed to engender independent local control loops managing the “business” of appliances, sensors, and actuators while still permitting occasional “advise and consent” communications with central servers.

XXV

INTRODUCTION

Finally, and most importantly, traditional IP peer-to-peer relationships lock out much of the potential richness of the Internet of Things. There will be vast streams of data flowing, many of which are unknown or unplanned. Only a publish/subscribe architecture allows us to tap into this knowledge by discovering interesting data flows and relationships. And only a publish/subscribe network can scale to the tremendous size of the coming Internet of Things.

The only systems on earth that have ever scaled to the size and scope of the Internet things are natural systems: pollen distribution, ant colonies, redwoods, and so on. From examining these natural systems, I developed the concept of a three-tiered IoT architecture described in this book: simple end devices; networking specialist propagator nodes, and information-seeking integrator functions. In these pages, lll explain why terse, self-classified messages, networking overhead isolated to a specialized tier of devices, and the publish/subscribe relationships formed are the only way to fully distill the power of the coming Internet of Things.

Francis daCosta Santa Clara, California, 2013

XXvi

CHAPTER 1 It’s Different Out Here

The emergence of the Internet of Things (IoT) destroys every precedent and preconceived notion of network architecture. To date, networks have been invented by engineers

skilled in protocols and routing theory. But the architecture of the Internet of Things will rely much more upon lessons derived from nature than traditional (and ossified, in my opinion) networking schemes. This chapter will consider the reasons why the architecture for the Internet of Things must incorporate a fundamentally different architecture from the traditional Internet, explore the technical and economic foundations of this new architecture, and finally begin to outline a solution to the problem.

Why the Internet of Things Requires a New Solution

The architecture of the original Internet was created long before communicating with billions of very simple devices such as sensors and appliances was ever envisioned.

The coming explosion of these much simpler devices creates tremendous challenges for the current networking paradigm in terms of the number of devices, unprecedented demands for low-cost connectivity, and impossibility of managing far-flung and diverse equipment. Although these challenges are becoming evident now, they will pose a greater, more severe problem as this revolution accelerates. This book describes a new paradigm for the Internet of Things; but first, the problem.

It’s Networking on the Frontier

The IoT architecture requires a much more organic approach compared with traditional networking because it represents an extreme frontier in communications. The scope and breadth of the devices to be connected are huge, and the connections to the edges of the network where these devices will be arrayed will be “low fidelity”: low-speed, lossy (where attenuation and interference may cause lost but generally insignificant data, as depicted in Figure 1-1), and intermittent. At the same time, much of the communication will be machine-to-machine and in tiny snatches of data, which is completely the opposite of networks such as the traditional Internet.

Caw! “| hear a crow”

Caw!

Caw!

Caw!

Beep!

Beep!

Beep! “My car Beep! Beep! alarm!”

Can you come

(( over here and “Is someone ~ help me? speaking?” ~

= - Attenuation, Interference, Received Signal Original Signal and Loss

Figure 1-1. The results of a lossy connection at an end point

Exploring the characteristics of the traditional Internet highlights the very different requirements for the frontier of the emerging Internet of Things. Conventionally, data networks have been over-provisioned; that is, built with more capacity than is typically required for the amount of information to be carried. Even the nominally “best effort” traditional Internet is massively over-provisioned in many aspects. If it weren’t, the Internet couldn’t work: protocols such as TCP/IP are fundamentally based on a mostly reliable connection between sender and receiver.

Because Moore’s Law provided a “safety valve” in the form of ever-increasing processor speeds and memory capacities, even the explosive growth of the Internet over the last two decades has not exceeded the capabilities of devices such as routers, switches, and PCs, in part because they are continually replaced at 3- to 5-year intervals with devices with more memory and processing power.

These devices are inherently multipurpose: they are designed with software, hardware, and (often) human access and controls. What is important about this point is that the addition of networking capability, usually in the form of protocol “stacks,” is nearly free. The processor power, memory, and so on already exist as byproducts of the devices’ prime functions.

But the vast majority of devices to be connected in the coming IoT are very different. They will be moisture sensors, valve controls, “smart dust,’ parking meters, home appliances, and so on. These types of end devices almost never contain the processors, memory, hard drives, and other features needed to run a protocol stack.

These components are not necessary for the end devices’ prime function, and the costs of provisioning them with these features would be prohibitive, or at least high enough to exclude wide use of many applications that could otherwise be well served. So these simpler devices are very much “on their own” at the frontier of the network.

Today’s Internet doesn’t reach this frontier; it simply isn’t cost-effective to do so, as will be explored later. Thus, it isn’t possible to overprovision in the same way networks have traditionally been built. On the frontier, devices in every aspect should therefore be more self-sufficient, from their naming, to protocols, to security. There simply isn’t the “safety net” of device performance, over-provisioning, a defined end-to-end connection, and management infrastructure as in traditional networking.

It Will be (Even) Bigger than Expected

As a growing number of observers realize, one of the most important aspects of the emerging Internet of Things is its incredible breadth and scope. Within a few years, devices on the IoT will vastly outnumber human beings on the planet—and the number of devices will continue to grow. Billions of devices worldwide will form a network unprecedented

in history. Devices as varied as soil moisture sensors, street lights, diesel generators, video surveillance systems—even the legendary Internet-enabled toasters—will all be connected in one fashion or another. See Figure 1-2 for some examples.

Figure 1-2. A wide variety of end devices will be connected to the Internet of Things

Some pundits have focused only on the myriad addresses necessary for the sheer arithmetic count of devices and have pronounced IPv6 sufficient for the IoT. But this mistakes address space for addressability. No central address repository or existing address translation scheme can possibly deal with the frontier aspects of the IoT. Nor can addresses alone create the costly needed networking “horsepower” within the appliances, sensors, and actuators.

Devices from millions of manufacturers based in hundreds of countries will appear on the IoT (and disappear) completely unpredictably. This creates one of the greatest challenges of the IoT: management. This is a matter both of scope and device capabilities.

Consider smartphones, for example, which are expected to become the most common computing and communications platforms in the world. This number has recently been placed at 1.4 billion, or roughly one for every five persons on the planet.

A similar figure has been estimated for PCs, bringing the total worldwide for these two types of devices to about 3 billion.

These devices incorporate the processors, memory, and human interfaces necessary for traditional networking protocol stacks (typically IPv6 today), the human interfaces necessary for control, and an infrastructure for management (unique addresses, management servers, and so on). The prices (and profit margins) of these devices mean that it is cost-effective for manufacturers (and governments) to keep track of addresses, feature sets, software revisions, and so on.

But the situation for the actuators, sensors, and appliances of the Internet of Things is vastly different. Considering the number of appliances per citizen in developed countries alone, the number is staggering: each of these individuals probably makes use of dozens of these devices each day. Even residents of developing countries interact with multiple end devices and sensors daily—and those numbers are growing with rising standards of living. Add to that a vast array of traffic-light controls, security devices, and status sensors operated by various levels of government, and the number of potential IoT end devices rapidly grows to a couple of orders of magnitude greater than the world’s population (7 billion and counting, as of this writing).

The estimated 700 billion IoT devices (see Figure 1-3) cannot be individually managed; they can only be accommodated. It will simply not be possible to administer the addressing of this huge population of communicating machines through traditional means such as IPv6 nor will it be necessary to do so. Instead, self-addressing and self-classification will provide the answers, as explained in Chapter 3.

Internet Core

Traditional Internet 3 billion to 5 billion devices

Internet of Things 700 billion to 1 trillion devices

Figure 1-3. The quantity of devices in the Internet of Things will dwarf the traditional Internet and thus cannot be networked with current protocols, tools, and techniques

Terse, Purposeful, and Uncritical

The kinds of information these hundreds of billions of IoT devices exchange will also be very different from the traditional Internet—at least the Internet we’ve known since the 1990s. Much of today’s Internet traffic is primarily human-to-machine oriented. Applications such as e-mail, web browsing, and video streaming consist of relatively large chunks of data generated by machines and consumed by humans. As such, they tend to be asymmetrical and bursty in data flows, with a relatively large amount of data exchanged in each “session” or “conversation.”

But the typical IoT data flow will be nearly diametrically opposed to this model. Machine-to-machine communications require minimal packaging and presentation overhead. For example, a moisture sensor in a farmer’s field may have only a single value to send of volumetric water content. It can be communicated in a few characters of data, perhaps with the addition of a location/identification tag. This value might change slowly throughout the day, but the frequency of meaningful updates will be low. Similar terse communication forms can be imagined for millions of other types of IoT sensors and devices. Many of these IoT devices may be simplex or nearly simplex in data flows, simply broadcasting a state or reading over and over while switched on without even the capacity to “listen” for a reply.

This raises another aspect of the typical IoT message: it’s individually unimportant. For simple sensors and state machines, the variations in conditions over time may be small. Thus, any individual transmission from the majority of IoT devices is likely completely uncritical. These messages are being collected and interpreted elsewhere in the network, and a gap in data will simply be ignored or extrapolated (see Figure 1-4).

Copious Data Reduced Data

Figure 1-4. Multiple identical messages may be received; some are discarded

Even more complex devices, such as a remotely monitored diesel generator, should generate little more traffic, again in terse formats unintelligible to humans, but gathered and interpreted by other devices in the IoT. Overall, the meaningful amount of data generated from each IoT device is vanishingly small—nearly exactly the opposite of the trends seen in the traditional Internet. For example, a temperature sensor might generate only a few hundred bytes of useful data per day, about the same as a couple of smartphone text messages. Because of this, very low bandwidth connections might be utilized for savings in cost, battery life, and other factors. On the IoT frontier, just as in the mythical “Old West,’ laconic characters will be appreciated.

Dealing with Loss

Today’s traditional Internet is extremely reliable, even if labeled “best effort.” Over-provisioning of bandwidth (for normal situations) and backbone routing diversity have created an expectation of high service levels among Internet users. “Cloud” architectures and the structure of modern business organizations are built on this expectation of Internet quality and reliability.

But at the extreme edges of the network that will make up the vast statistical majority of the IoT, connections may often be intermittent and inconsistent in quality. Devices may be switched off at times or powered by solar cells with limited battery back-up. Wireless connections may be of low bandwidth or shared among multiple devices.

Traditional protocols such as TCP/IP are designed to deal with lossy and inconsistent connections by resending data. Even though the data flowing to or from any individual IoT device may be exceedingly small, it will grow quite large in aggregate IoT traffic. The inefficiencies of resending vast quantities of mostly individually unimportant data are clearly an unnecessary redundancy. Again, recall that for the vast majority of IoT devices, a lost message (or even a substantial string of messages) is not meaningful.

(For those devices that are sending or receiving timely mission-critical information, traditional Internet protocols are likely a better fit than the emerging IoT architecture.)

The Protocol Trap

It’s extremely tempting to suggest existing widely deployed protocols such as TCP/IP

for the IoT (see the sidebar “ Why not IP for the IoT?” in Chapter 2). After all, they have already been engineered and are widely available in protocol stacks on billions of devices such as PCs and smartphones. But, as briefly noted, most of these protocols are ill-suited for many of the end devices with potential interest for the IoT.

The basic problem is the very robustness of these protocols. They are intrinsically designed for high-duty cycles, large data streams, and reliability. Each of these otherwise desirable characteristics is a poor fit for the IoT, as noted previously. But what’s the harm, one might ask? Isn’t more capability a good thing? Not for the Internet of Things.

Mind the Overhead

A key reason why robust protocols aren’t needed (or possible) for the IoT is the overhead they require and the minimal processing, memory, and communications capabilities

of many very simple IoT devices. This may come as a shock to some IoT thinkers who envision an IP stack on every light post and refrigerator. But when the IoT is considered from the proper “end of the telescope” —from the edge of the network in—this immediately becomes impractical, for all the reasons noted previously. Instead, it makes sense to provide a new solution that can run side by side with existing IP-enabled end devices to efficiently manage the immense amount of data being generated by devices for which IP support is unnecessary and perhaps a liability.

Much of what has been written to date about the IoT assumes a sophisticated networking stack in every refrigerator, parking meter, and fluid valve, so this may be a difficult idea to abandon. But from the forgoing discussion, it’s obvious that these devices won't need the decades of built-up network protocol detritus encoded in TCP/IP, for example. One must free his or her thinking from personal experiences and concepts of the networking of computers, smartphones (and, by definition, human users) to address the much simpler needs of the myriad devices at the edge of the IoT.

Burdening otherwise simple devices such as power line sensors and coffee makers with a full networking protocol stack would serve only to massively increase the cost and complexity of billions of these devices. A traditional networking protocol stack requires a processor, operating system, memory, and other functions. Even if consolidated within a single chip, the complexity, power draw, and cost of this computing power is an unnecessary expense in the IoT. These costs will be considered later in this chapter.

As noted previously, the vast majority of IoT devices have very basic needs of sending or receiving a miniscule amount of data. The physical requirements may likewise be very simple: an integrated chip containing only the minimal interfaces and a means of transmission or reception.

More Smarts, More Risk

Although it may seem counterintuitive, dumber devices are safer. If every IoT device has some sort of operating system and memory, it becomes a potential subject for hacking

or inadvertent misconfiguration. The operating systems and protocol stacks also require updating and management. Providing security and upgrades on the scale of the IoT for a massive number of devices, built and installed by millions of different manufacturers and individuals, is simply an impossible task (see Figure 1-5).

MAC/PHY (radio, et al)

Receiver

Memory

Operating System

Processor PHY only (radio, et al) Transmitter or Receiver

Modulator

TO Logie

4 yyy

IP-enabled device loT protocol device

Figure 1-5. Contrasting the processor, OS, memory, and power necessary for traditional protocols vs. the IoT protocol

The Overhead of Overhead

Beyond the physical costs and management requirements, the data overhead of traditional networking is likewise overkill for the majority of the IoT. Traditional protocols are “sender-oriented”; that is, the sender must ensure that its message has been properly transmitted and received. This leads to extensive capabilities in terms of temporary storage of sent data, management of acknowledgments, and resending of lost or corrupted messages. And each of these robust capabilities is reflected in overhead data added to the message payload.

When this data overhead is considered in relation to the tiny snatches of data sent or received by the typical IoT device, the ratio of overhead to payload becomes ridiculous. Moreover, because each individual IoT message is completely uncritical, the check-andretransmit overhead is an unnecessary expense in bandwidth and end device cost. It makes the most sense, therefore, for the emerging IoT architecture to be engineered for an absolute minimum of data overhead.

Humans Need Not Apply

Perhaps most importantly, traditional networking protocols and applications are almost all designed with the expectation of a human being on one end of the “conversation.” These traditional approaches are inherently designed to communicate concepts and context for humans.

But the networking overhead associated with smooth streaming, echoing of typed characters, and intelligible presentation of data are completely unnecessary at the machine-to-machine device level in the Internet of Things. So a large percentage of the processing and data overhead of traditional protocols is totally redundant for the IoT. An architecture for the Internet of Things should provide only the minimal amount of overhead that is needed—and only at the point that it is needed—to maximize efficiency and minimize costs.

Economics and Technology of the Internet of Things

One of the great promises of bringing IPv6 to the traditional Internet was that it would provide all the address space needed to connect every device ever needed forever— including the Internet of Things, no matter how large it grew. And within that narrow definition, the promise is correct. Because of some quirks in the way that only part

of the IPv6 address space has been released, the current theoretical number of hosts (communicating devices) on an IPv6 Internet is 3.4x10*38*.

This is indeed a huge number, which even the massive Internet of Things is unlikely to surpass. For this reason, many pundits and manufacturers (particularly those with a vested interest) have sanguinely said that IPv6 is already prepared for the Internet of Things. The world simply needs to keep doing what it has always done to incorporate the new IJoT—there are more IP addresses available than grains of sand.

But this “head in the sand” approach ignores the key economic factor that will drive the deployment of the Internet of Things (as it has driven nearly every other networking technology): the cost at the end points. There are three broad areas where these costs accumulate and compel the need for a new approach in the Internet of Things: hardware and software, oversight and management, and security.

Functionality Costs Money

As noted earlier, traditional computing and communications devices such as PCs, tablets, and smartphones already incorporate processors, working memory, and storage in their design. These capabilities are necessary for their primary purpose. Adding IPv6 to these devices requires only the addition of a protocol stack that resides in storage, executes within working memory, and is powered by the processor.

Thus the incremental cost of adding IPv6 to these devices is indeed negligible, in fact barely measurable, when compared with the profit margins these devices generate. But these devices are not a significant portion of the Internet of Things! Numbering in the low billions today, their number will be dwarfed by the hundreds of billions of simple sensors and appliances in the IoT.

The vast majority of these simple end devices contain no processors, memory, or storage; and are not data-connected in any way today. This is a key point: the future of the Internet of Things is networking devices that have never been connected before. These devices are designed to be built and sold, for the most part, at the lowest cost yielding the highest margin. Those sold in developing countries, in particular, must be extremely inexpensive. Yet they are some of the very areas in which the IoT will grow most quickly. To capitalize on the enormous potential of the IoT, creating a standard low-cost solution will enable billions of devices that would otherwise continue to be off the grid, never developed, or added to the massive quantity of one-off solutions that are being spawned even today.

Inexpensive Devices Can’t Bear Traditional Protocols

With a clearer picture of these cost realities in mind, it is immediately obvious that burdening moisture sensors, light bulbs, and the proverbial toaster with the additional hardware and software (not necessary for the basic functions of these sensors and appliances) needed to run traditional protocols such as IPv6 is a show-stopper. It has been estimated that the incremental cost of adding IPv6 to devices can be as much as $50, even in large quantities. (Note that beyond the processors and memory devices, additional Wi-Fi or Ethernet components are needed, and more power and heat dissipation will also be required).

Fortunately for the expansion of the Internet of Things, these simple devices do not require anything approaching the level of complexity offered by IPv6. Instead, simple modulation, broadcast, and receiving technologies will suffice, even including non-radio-frequency solutions such as infrared and power line networking. Assuming integration into silicon packages, costs for adding simple IoT networking (described in Chapter 2) to sensors and appliances will quickly approach $1 or less. The key is that this is barely “networking” in the traditional sense: broadcasting a state or receiving a simple instruction with no error correction, routing, or any other traditional networking functions. IoT devices are “dumb” in general, but they are exceedingly well-suited to a narrow task. At a very base level, it is easy to see that this cost argument alone is proof that the costs and the effort in creating a new solution for IoT devices are absolutely necessary. The result in not doing so would be that many of these new technologies and innovations would largely not come to pass. Others would be implemented at a cost that limits their usefulness. At what cost to growth, development, and prosperity?

And as noted previously, traditional one-size-fits-all networking protocols such as IPv6 burden even the smallest payloads with 1,000 bytes of data. In today’s over-provisioned world, these wasted bytes are unnoticed. But when extrapolated to hundreds of billions of simple end devices sending and receiving hundreds of thousands of times each day, the potential for network congestion and huge expenditures by carriers is significant. New carrier build-outs to support the “plain vanilla” data networking of the IoT will be difficult to cost-justify.

Overseeing 700 Billion Devices

The count of manufacturers building networking equipment likely numbers in the millions. They are relatively easy to find and track because each traditional piece of networking equipment is associated with a MAC ID (Media Access Control Identification) assigned to the manufacturer. A large number, but there is a central database of manufacturers that is maintained by the IEEE (Institute of Electrical and Electronics Engineers).

For those manufacturers who are today building traditional networking equipment, one may assume a significant amount of networking knowledge. Imagine the impact ofa new IoT standard on the number of network-ready manufacturers out there and the boost that would give to the worldwide economy.

Contrast this with the likely millions of firms and individuals worldwide building the kinds of simple sensors, actuators, and appliances which will be connected to the Internet of Things. It is inconceivable that all those makers of simple devices can be expected to queue up for addresses assigned by any centralized authority—or that rogue states, organizations, or individuals wouldn’t attempt to subvert such systems.

Extending this thinking, simply scanning for hundreds of billions of IPv6 addresses would take literally hundreds of years. It is one thing to put addresses on nearly a trillion devices, but quite another to find and manage one device out of that constellation. The human cost to manage an Internet of Things made up solely of sophisticated IPv6 devices would exceed the cost of any networking project on earth to date. These costs will fall hardest on already strapped carriers that are already struggling to wring more revenue from expensive physical plant investments.

Only Where and When Needed

Of necessity, the emerging new architecture of the Internet of Things should take an entirely different approach, as described throughout this book. End devices have only locally meaningful and likely non-unique names. This is not a problem because there is networking intelligence elsewhere in the architecture at a much smaller (and thus more manageable) number of points.

And there is no need to oversee or control every maker of end devices. Because the IoT provides only limited networking capabilities at the end devices, there is little “harm” they can do on the network as a whole, and this is easily controlled through a much smaller number of “smarter” devices.”

This approach is totally different from IPv6, which demands that every device have the functionality and management to act as a “peer” on the network. The Internet of Things simply cannot scale if built of peers that all must be managed. Like a massive ant colony, the IoT will scale through specialization, individual autonomy, and localized effect. In this way, costs are reduced by orders of magnitude.

Security Through Simplicity (and Stupidity)

A trite statement, but ultimately true. Because the communications with the end devices in this emerging architecture of the Internet of Things are so basic and so specialized, there are limited back doors and security risks. Again, contrast this with the “peer-to-peer” world of the IPv6 Internet where many IP devices are exposed to hacking and ll cracking attempts from anywhere in the world. The global cost of Internet security breaches has been estimated at $115 billion (Symantec, 2012). With roughly 2.4 billion peer-to-peer nodes on the Internet today, this roughly equates to $50 per node (user) per year in losses. Multiplying that figure times the projected hundreds of billions of Internet of Things devices creates an unsustainably high cost of IPv6 in the IoT.

By focusing on limited networking capabilities for the end devices as described in this book, the emerging architecture of the Internet of Things drastically reduces the risks and costs associated with networking the huge population of appliances, actuators, and sensors.

Cost and Connectivity

The key for the expected expansion of the Internet of Things is connecting hundreds of billions more devices at far-reduced costs and risks. Only this emerging IoT architecture can accomplish both in a way that is cost-effective for device manufacturers, Internet carriers, and users.