7.3: Broader Perspective

Last updated
Save as PDF

Page ID: 13983

\( \newcommand{\vecs}[1]{\overset { \scriptstyle \rightharpoonup} {\mathbf{#1}} } \) \( \newcommand{\vecd}[1]{\overset{-\!-\!\rightharpoonup}{\vphantom{a}\smash {#1}}} \)\(\newcommand{\id}{\mathrm{id}}\) \( \newcommand{\Span}{\mathrm{span}}\) \( \newcommand{\kernel}{\mathrm{null}\,}\) \( \newcommand{\range}{\mathrm{range}\,}\) \( \newcommand{\RealPart}{\mathrm{Re}}\) \( \newcommand{\ImaginaryPart}{\mathrm{Im}}\) \( \newcommand{\Argument}{\mathrm{Arg}}\) \( \newcommand{\norm}[1]{\| #1 \|}\) \( \newcommand{\inner}[2]{\langle #1, #2 \rangle}\) \( \newcommand{\Span}{\mathrm{span}}\) \(\newcommand{\id}{\mathrm{id}}\) \( \newcommand{\Span}{\mathrm{span}}\) \( \newcommand{\kernel}{\mathrm{null}\,}\) \( \newcommand{\range}{\mathrm{range}\,}\) \( \newcommand{\RealPart}{\mathrm{Re}}\) \( \newcommand{\ImaginaryPart}{\mathrm{Im}}\) \( \newcommand{\Argument}{\mathrm{Arg}}\) \( \newcommand{\norm}[1]{\| #1 \|}\) \( \newcommand{\inner}[2]{\langle #1, #2 \rangle}\) \( \newcommand{\Span}{\mathrm{span}}\)\(\newcommand{\AA}{\unicode[.8,0]{x212B}}\)

7.3 Broader Perspective

Big Data and Analytics

This chapter is about data, and since no topic in Computer Science is receiving more attention than big data (or alternatively, data analytics), a natural question is what relationship there might be between big data and computer networks. Although the term is often used informally by the popular press, a working definition is quite simple: sensor data is collected by monitoring some physical or man-made system and then analyzed for insights using the statistical methods of Machine Learning. Because the amount of raw data that’s collected is often voluminous, the “big” qualifier is applied. So are there any implications to networking?

At first blush, networks are purposely designed to be data-agnostic. If you collect it and want to ship to somewhere for analysis, the network is happy to do that for you. You might compress the data to reduce the bandwidth required to transmit it, but otherwise big data is no different than plain old regular data. But this ignores two important factors.

The first is that while the network doesn’t care about the meaning of the data (i.e., what the bits represent), it does concern itself with the volume of data. This impacts the access network in particular, which has been engineered to favor download speeds over upload speeds. That bias makes sense when the dominant use case is video that flows out to end-users, but in a world where your car, every appliance in your house, and the drones flying over your city are all reporting data back into the network (uploaded into the cloud), the situation is reversed. In fact, the amount of data being generated by Autonomous Vehicles and the Internet-of-Things (IoT) is potentially overwhelming.

While one could imagine dealing with this problem by using one of the compression algorithms described in Section 7.2, people are instead thinking outside the box, and pursuing new applications that reside at the edge of the network. These edge-native applications both provide better sub-millisecond response time and they dramatically reduce the volume of data that ultimately needs to be uploaded into the cloud. You can think of this data reduction as application-specific compression, but it’s more accurate to say that the edge application needs only write summaries of the data, not the raw data, back to the cloud.

We introduced the access-edge cloud technology needed to support edge-native applications at the end of Chapter 2, but what is perhaps more interesting is to look at some examples of edge-native applications. One such example is enterprises in the automotive, factory, and warehouse space increasingly want to deploy private 5G networks for a variety of physical automation use cases. These include a garage where a remote valet parks your car or a factory floor making use of automation robots. The common theme is high bandwidth, low latency connectivity from the robot to intelligence sitting nearby in an edge cloud. This drives lower robot costs (you don’t need to place heavy compute on each one) and enables robot swarms and coordination more scalably.

Another illustrative example is Wearable Cognitive Assistance. The idea is to generalize what navigation software does for us: it uses one sensor (GPS), gives us step-by-step guidance on a complex task (getting around an unknown city), catches our errors promptly, and helps us recover. Can we generalize this metaphor? Could a person wearing a device (e.g., Google Glass, Microsoft Hololens) be guided step-by-step on a complex task, perhaps for the first time? The system would effectively act as “an angel on your shoulder.” All the sensors on the device (e.g., video, audio, accelerometer, gyroscope) are streamed over wireless (possibly after some device preprocessing) to a nearby edge-cloud that performs the heavy lifting. This is a human-in-the-loop metaphor, with the “look and feel of augmented reality” but implemented by AI algorithms (e.g., computer vision, natural language recognition.)

The second factor is that because a network is like many other man-made systems, it is possible to collect data about its behavior (e.g., performance, failures, traffic patterns), apply analytics programs to that data, and use the insights gained to improve the network. It should not come as a surprise that this is an active area of research, with the goal of building a closed control loop. Setting aside the analytics itself, which are well outside the scope of this book, the interesting questions are (1) what useful data can we collect, and (2) what aspects of the network are most promising to control? Let’s look at two promising answers.

One is 5G cellular networks, which are inherently complex. They include multiple layers of virtual functions, virtual and physical RAN assets, spectrum usage, and as we have just discussed, edge computing nodes. It is widely expected that network analytics will be essential to building a flexible 5G network. This will include network planning, which will need to decide where to scale specific network functions and application services based on machine learning algorithms that analyze network utilization and traffic data patterns.

A second is In-band Network Telemetry (INT), a framework to collect and report network state, directly in the data plane. This is in contrast to the conventional reporting done by the network control plane, as typified by the example systems described in Section 9.3. In the INT architecture, packets contain header fields that are interpreted as “telemetry instructions” by network devices. These instructions tell an INT-capable device what state to collect and write into the packet as it transits the network. INT traffic sources (e.g., applications, end-host networking stacks, VM hypervisors) can embed the instructions either in normal data packets or in special probe packets. Similarly, INT traffic sinks retrieve (and optionally report) the collected results of these instructions, allowing the traffic sinks to monitor the exact data plane state that the packets “observed” while being forwarded. INT is still early-stage, and takes advantage of the programmable pipelines described in Section 3.4, but it has the potential to provide a qualitatively deeper insights into traffic patterns and the root causes of network failures.

Broader Perspective

To continue reading about the cloudification of the Internet, see Blockchain and a Decentralized Internet.

To learn more about promising edge-native applications, we recommend:

Open Edge Computing Initiative, 2019.

To learn more about In-band Network Telemetry, we recommend:

In-band Network Telemetry via Programmable Dataplanes, August 2015.