The sections so far have described how networks are built and how data moves through them. This section steps back to ask what affects a network once it exists and people depend on it. It first sets out the measures used to describe how well a network performs, since the effect of any factor shows up in them. It then takes the factors in five groups, which differ in kind as well as in subject. Scale is a matter of demand: how many people and devices use the network, and when. Access is a matter of geography and economics, of who can reach a network at all. Reliability concerns physical things, such as cables, power, and weather. Security and user behavior bring in the people, whose habits affect both the load on a network and its risks, and the section ends with one example where convenience and security pull directly against each other, account recovery.

Measuring performance#

People often describe a connection by a single number, its speed, but several distinct measures are involved. Bandwidth is a connection's maximum data capacity, the figure an Internet plan advertises, such as 100 megabits per second. Throughput is the rate actually achieved, which can be much lower. A speed test that reports 40 megabits per second on a 100-megabit plan is measuring throughput, while the 100 is the bandwidth. Performance also depends on the whole path, not on any one link. Data crossing several links can move no faster than the slowest of them, and other factors such as latency matter too, so if a video call to another school is poor even after a school upgrades its own connection, another part of the path, such as a slower link, is probably limiting it.

Three further measures describe the timing of data, not its quantity. Latency is the delay in data travel, the time for data to get from one end to the other. Jitter is the variation in that delay. Two meetings can have the same average delay, but if packets in one of them arrive at uneven intervals, that one has jitter. Packet loss is data that never arrives. Each of these especially harms real-time applications, which must play data at the moment it is due, such as a live video conversation with a classroom in another country. A half-second delay adds almost nothing to a bulk file transfer or an email message, but it makes a conversation awkward. For recorded video, jitter is smoothed by buffering: the player stores a few seconds of video ahead and plays from that store at a steady pace, even when packets arrive unevenly. A live call can't buffer much without adding delay, and as 0003.e explained, real-time applications often skip retransmission, because a resent packet would arrive too late to be played at the right moment, and a small gap is less disruptive than a growing delay.

These measures also reveal when a network is running out of capacity. As a network grows, falling available bandwidth and rising latency are key signs that it is approaching its limit. When throughput nears the bandwidth of a link, traffic competes for the remaining capacity and users experience slower responses. A server whose throughput stops rising as demand grows, while its response times climb, has hit a capacity limit somewhere in the system.

Scale: congestion and growth#

The most common cause of reaching that limit is simply sharing. Congestion occurs when many users share limited capacity, as when the residents of a neighborhood together demand more than their shared link can carry. This is why networks often slow at peak times, such as evenings at home or the first minutes of a class period at school. The packet switching described in 0003.e is what makes sharing economical, and congestion is its cost. TCP responds automatically: packet loss is a sign of congestion, so a TCP sender resends the missing data and reduces its sending rate. People can respond too, by spreading demand out. If every class streams the same video over a school's Internet link, storing a copy on a local server, which is a form of caching, means classes no longer use the link for it, and scheduling viewings at different times has a similar effect. Adding capacity helps, but often for less time than expected, because demand tends to grow into new capacity: after an upgrade, residents' use of data can grow to fill it, and congestion returns even though the number of users stays the same.

Growth in the number of connected devices raises the stakes. The Internet of Things (IoT) refers to everyday devices that connect to networks, such as thermostats, cameras, door locks, and sensors. Each device needs a network address to communicate, so their growing numbers increase the demand for addresses, one reason for the move to IPv6 described in 0003.e. They also increase traffic, since small amounts of data from a great many devices add up. And they increase the number of potential points of attack, particularly because many devices are rarely updated, so a flaw discovered in one may stay open for as long as it stays connected. A common response is to place such devices on a separate network segment, apart from sensitive systems such as student records, which limits what a compromised device can reach.

Access: geography, cost, and new technology#

Scale is a problem of having many users, and access is the problem of reaching users at all. Network access depends on infrastructure cost and population density. Building a network means laying cable or putting up towers, and where homes are far apart, the cost per customer is much higher, so providers see too little revenue to justify building. This is why rural communities are often underserved. Low-income communities are often underserved as well, even where lines exist, because the availability of a service doesn't by itself make it affordable. Two kinds of response appear in practice. Public funding can cover part of the cost of building networks where that cost exceeds expected revenue. Cooperatives, networks owned by their members, can serve areas where providers see too little profit, because their decisions can reflect local needs instead of outside profit targets. Section 0001.c places these gaps within the wider digital divide.

New wireless technologies are sometimes presented as the answer to these gaps, and their physics shows why the answer is mixed. High-frequency 5G cellular service offers high speeds but short range, so it needs many cell sites. Many sites are most economical where many customers live close together, so 5G tends to favor dense, already well-served areas, and a new technology doesn't automatically reach underserved communities. Satellites take a different approach. A traditional geostationary satellite orbits far from Earth, staying over one spot, and the long distance its signals travel adds latency. Low-Earth-orbit satellites are much closer to Earth, so signals travel a much shorter distance and latency is lower, and they can bring service to remote regions without building wired networks or many cell sites. Because each low satellite covers a small area and moves quickly across it, continuous coverage requires many of them.

Reliability: physical factors#

A network that people can reach still has to keep working, and reliability depends on factors that are largely physical: redundant paths, equipment quality, power, weather, and physical damage to cables. Storms, power failures, and severed cables are typical causes of outages. Each factor has a matching protection. An uninterruptible power supply (UPS), a battery unit between the equipment and the outlet, bridges short outages so equipment keeps running during brief power failures. A second connection along a different physical path protects against a cable being cut, but only if the paths are truly independent, since damage to a shared trench affects every cable in it, the same lesson about shared risks that 0003.d drew for redundancy in general. Equipment quality matters most where an outage would have serious consequences, so a router that carries critical services deserves better equipment than one serving a break room.

Security and user behavior#

The factors so far are largely technical, but networks are used by people, and user behavior affects both the load on a network and its security. On the load side, streaming and peak use add traffic, as when every student logs in to the same service at the start of the hour. On the security side, weak passwords and clicks on malicious links introduce risks, as when someone follows a link in an unexpected email and signs in on a fake page, giving an attacker a way in. Routine use such as streaming mainly adds load, while a single careless click can compromise an account. Because both problems come from behavior, the remedies address behavior. Staggering class start times spreads logins across the hour and reduces the peak a server must handle. Scheduling updates after hours moves their load away from the school day, and training staff to recognize fraudulent messages addresses phishing. Policies need to account for how people will actually behave, since burdensome rules invite workarounds, as a rule requiring a password change every week tends to produce passwords written on sticky notes.

Account recovery shows the tension between convenience and security especially clearly. Many sites let a user who forgets a password receive a reset link by email. Email-based password resets make accounts easier to recover, but they reduce security, because the email account becomes a single point of compromise: anyone who controls a user's mailbox can request reset links and take over the accounts tied to it, without knowing any of their passwords. Several measures limit the risk. Requiring a second factor, such as a code from the user's phone along with the emailed link, protects the account even if the email has been compromised. Short-lived, single-use links reduce the chance that a stolen or forwarded link is used later. Security questions are a weak alternative, because their answers, such as a pet's name or a hometown, can often be found or guessed from public sources such as social media. All of these factors act on data as it moves through a network, and underneath every one of them the data is just bits, which is the subject of the final section of this part.

• bandwidth: a connection's maximum data capacity
• throughput: the rate of data transfer actually achieved
• latency: the delay in data travel
• jitter: variation in latency
• packet loss: data that never arrives
• buffering: storing data ahead so that playback can proceed at a steady pace
• congestion: slowdown when many users share limited capacity
• caching: keeping a nearby copy of data to avoid fetching it again
• Internet of Things (IoT): everyday devices connected to networks
• network segment: a separate part of a network that limits what devices on it can reach
• cooperative: a network owned by its members
• 5G: high-frequency cellular service with high speeds and short range
• low-Earth-orbit satellite: a satellite close to Earth, giving lower latency than a geostationary one but covering a small, moving area
• uninterruptible power supply (UPS): a battery unit that bridges short power outages
• email-based password reset: account recovery through a link sent by email
• security questions: personal questions used to verify identity, weak because answers can be found or guessed

Practice: practice questions for 0003.g