Matheion

MA-NET1 · Module 3

Delay, loss, and throughput

Topics: Transmission delay · Propagation delay · Queuing · Bottleneck throughput

Overview

Session 3 — Delay, loss, and throughput

Mira notices that her forecast sometimes appears instantly and sometimes drags, even though nothing about her phone has changed. In this session we give her the vocabulary and the arithmetic to explain why. We separate delay into four distinct causes, learn which of them stay the same from packet to packet and which wobble, work out the total time for a packet to cross several store-and-forward links, and then switch viewpoint to throughput, the rate at which data actually arrives, and the bottleneck that limits it.

By the end you should be able to: name and compute transmission delay (L/R) and propagation delay (d/s); recognise nodal processing and queuing delay and when each matters; add up end-to-end delay across multiple links; find a path's bottleneck; and use traffic intensity as an intuition for when queues grow.

Scroll Concepts → Examples → Quiz. This is the most numerical session of the opening four, so plan on roughly 60–70 minutes and let yourself redo each calculation once on paper.

Tip: if a paragraph feels dense, pause and say it back in your own words before moving on; that small habit saves time later.

Concepts

Core ideas — where time goes: delay, loss, and throughput

Mira taps refresh on her Skylark Forecast app. Sometimes the new outlook appears before she lifts her finger; sometimes the spinner turns for several seconds. Her phone has not changed, the forecast has not changed, and yet the experience has. The difference lies in how long the data spends travelling and waiting along the path between her phone and api.skylark.example. In this session we learn to break that time into its distinct causes, to calculate the ones that can be calculated, and to understand the second big measure of network quality, throughput. All numbers and names are invented for practice.


One packet, one hop: where can time go?

Start small. Picture a single packet that has just arrived at a router and is about to leave on an outgoing link toward the next router. How much time does the packet spend at this router, and on the link beyond it?

It turns out that the time splits into four separate parts, and keeping them apart in your head is the single most useful skill in this session. They are:

  1. Nodal processing delay: time the router spends looking at the packet's header and deciding what to do with it.
  2. Queuing delay: time the packet spends waiting in the buffer before it can be sent.
  3. Transmission delay: time needed to push all of the packet's bits onto the link.
  4. Propagation delay: time for a bit, once on the link, to travel to the other end.

Add the four together and you have the total nodal delay at this hop. Let us meet each in turn, slowly, because the last two are often confused with each other.


Nodal processing delay

When a packet arrives, the router must read its header, check it for bit errors, and consult its forwarding table to choose an outgoing link. All of that takes a little time. In modern high-speed routers, this is typically on the order of microseconds, which is a thousandth of a millisecond. Because it is so tiny compared to the other delays on most paths, we will usually treat it as a small constant, or ignore it when a problem says to. It is mostly the same from one packet to the next, because the router does roughly the same work each time.


Queuing delay

After the router decides where the packet goes, the packet may have to wait for its turn on the outgoing link. If the link is idle and no other packet is waiting, the queuing delay is zero. If ten other packets are ahead of it, the packet must wait until all ten have been transmitted.

This is the one delay that varies from packet to packet. Two packets sent a moment apart by Mira's phone might find very different queues: the first slips through an empty router, while the second arrives just as a burst from another user has filled the buffer. Because of this variability, queuing delay is usually described by averages, spreads, or the chance that it exceeds some value, rather than by a single number.

It is worth saying plainly: queuing delay depends on traffic, not on distance or link design. It can be nothing at all or it can be large, and when it is large it tends to dominate the other three delays.


Transmission delay: pushing the bits onto the wire

Here is the delay beginners find most confusing, so we will take it step by step.

A link has a transmission rate (R), measured in bits per second. This is the speed at which the sender can put bits onto the link, one after another, in a continuous stream. A packet has a length (L), measured in bits. To get the whole packet onto the link, the sender must push out (L) bits at a rate of (R) bits per second, so the time required is

[ d_{\text{trans}} = \frac{L}{R}. ]

Think of a bathtub filling through a tap. The packet is the amount of water ((L)), and the tap's flow rate is (R). The time to pour the entire amount into the tub is the amount divided by the flow rate. For instance, a packet of 12,000 bits on a link of 4 Mbps takes 12,000 ÷ 4,000,000 = 0.003 s = 3 ms. A larger packet or a slower link means a longer wait; a smaller packet or faster link means a shorter one.

The unit conversions deserve care. Packet sizes are often given in bytes; to use the formula, multiply by 8 to get bits. Rates given in Mbps must be multiplied by 1,000,000 to get bps. Most calculation mistakes in this topic are unit mistakes, not conceptual ones.


Propagation delay: the bit's journey along the link

Once the first bit of a packet is on the link, it must physically travel to the other end. This travel time is the propagation delay. It depends on the length of the link, (d) (in metres), and the propagation speed of the signal in the medium, (s) (in metres per second). The speed is close to, but a little less than, the speed of light: for fiber and many copper cables it is roughly (2 \times 10^8) m/s. The formula is

[ d_{\text{prop}} = \frac{d}{s}. ]

For a link 600 km long, (d = 600{,}000) m, so with (s = 2 \times 10^8) m/s the propagation delay is 600,000 ÷ 200,000,000 = 0.003 s = 3 ms. Notice that the formula contains no mention of the packet's size or the link's rate. It depends purely on distance and the medium.


Transmission versus propagation: the picture that clears the fog

The two delays sound alike, and many learners mix them up. Here is an image to keep them apart.

Picture a bakery that sends trays of rolls to a café across town by conveyor belt. A worker stands at the start of the belt and places the rolls on it one by one. The belt then carries each roll, at a fixed walking-pace speed, to the café at the far end.

  • The time the worker needs to place a whole tray onto the belt is like transmission delay. It depends on how many rolls are on the tray (the packet's length) and how quickly the worker loads them (the link's rate).
  • The time a single roll spends riding the belt to the café is like propagation delay. It depends on the belt's length and its speed. It does not matter how many rolls are in the tray, nor how fast the worker loaded them.

Now see how the two respond to changes. If a quicker worker takes over, the tray is loaded sooner, but each roll still rides the belt for the same time. In networking terms: raising the rate (R) shrinks transmission delay but leaves propagation delay untouched. Conversely, if the belt is shortened, each roll arrives sooner, but loading the tray takes just as long as before. Distance affects propagation, not transmission.

Common confusion: "A faster link makes bits travel faster." It does not. A faster link puts bits onto the wire at a higher rate; it does not make each bit move down the wire faster. A bit on a 10 Mbps cable and a bit on a 10 Gbps cable (of the same material) both travel at about (2 \times 10^8) m/s.


Which delays are steady and which wobble?

We can now sort the four delays by how much they change from packet to packet.

  • Transmission delay (L/R): constant for a given packet length and link. Fixed.
  • Propagation delay (d/s): constant for a given link. Fixed, unless the route changes.
  • Nodal processing delay: small and nearly constant.
  • Queuing delay: variable, driven by traffic.

So when Mira sees her forecast time jump between 80 ms on one refresh and 600 ms on the next, the suspect is almost always queuing. The other three delays would give the same value each time unless the route or the packet sizes changed.


Adding up the delay along a whole path

Real packets cross several links. To compute the delay from a source host to a destination host, we add up the nodal delay at each hop. For a path with (N) links, so (N - 1) routers in between, the total is

[ d_{\text{end-to-end}} = \sum_{i=1}^{N} \left( d_{\text{proc},i} + d_{\text{queue},i} + \frac{L}{R_i} + \frac{d_i}{s} \right), ]

where link (i) has rate (R_i), length (d_i), and so on, and the end host that sends the packet plays the role of "first hop" for the transmission term. Do not be intimidated by the sum sign. It says: for each link, write the four delays and add; then add up all the links. When the routers are idle (no queuing) and processing is negligible, the formula reduces to the sum of (L/R_i) and (d_i/s) over the links.

Because of store-and-forward, each router has to receive the whole packet before sending it on, so the transmission term (L/R_i) appears once per link. A path of three equal links gives three transmission delays, not one. This is the detail that most often goes missing in student attempts, so burn it in: one transmission delay per link, one propagation delay per link.


Why sending in smaller packets can help

Suppose Mira's phone has a message of 100,000 bits to send over a path of several links. If it sends the message as a single huge packet, each router must receive all 100,000 bits before passing it on. The links further down sit idle while the first one is busy.

If instead the phone cuts the message into, say, ten packets of 10,000 bits, the first packet can already be moving on the second link while the second packet is still being sent on the first link. The links work in an overlapping, pipelined fashion, and the message arrives sooner overall. In general, for (K) links of equal rate and a message split into (P) equal packets, the time to deliver the whole message (counting only transmission) is ((P + K - 1)) packet-transmission-times, compared to (K \times P) packet-times if all links had to finish a packet before the next one began. The worked examples carry out this arithmetic. Splitting also helps for a second reason: a lost small packet is cheaper to resend than a lost huge one.


Traffic intensity: when queues start to grow

Queuing delay is not predictable from a formula in the way (L/R) is, but we can build a trustworthy intuition using one quantity. Let (a) be the average rate at which packets arrive at a link, in packets per second, and let (L) be the average packet length in bits. Then the average rate at which bits arrive is (L \cdot a) bits per second. Compare that to the rate (R) at which the link can send. The ratio

[ \text{traffic intensity} = \frac{L \cdot a}{R} ]

tells you how heavily the link is loaded. Read it as "average bits arriving per second divided by bits the link can serve per second".

What happens as this ratio changes?

  • Close to 0 (say 0.1): packets arrive rarely compared to how fast the link serves them. Most packets find the link idle, and queuing is minimal.
  • Moderate (say 0.5 to 0.8): the link is busy a good fraction of the time. Packets sometimes arrive in bunches and wait. Average queuing delay is noticeable and grows steadily.
  • Approaching 1: arrivals almost match capacity. Even a brief burst creates a queue that takes a long time to drain. The average delay grows very rapidly: small increases in load cause large increases in waiting.
  • Greater than 1: bits arrive faster than the link can send them. The queue grows without bound until the buffer fills and packets are lost.

The most important feature is that queuing delay is not linear in load. Going from 0.5 to 0.6 barely matters; going from 0.9 to 0.99 is dramatic. This is why network operators try to keep links comfortably below full.

Note also that traffic intensity depends on averages. Two links with the same average intensity can behave differently: if the arrivals are perfectly evenly spaced, queues stay tiny, but if packets arrive in tight bursts, queues build even when the average load is modest.


Packet loss, revisited

The last session explained that a packet is dropped when it finds the buffer full. Combined with the traffic-intensity picture, the story becomes tidy: the higher the intensity, the more likely the buffer is full, and the more packets are lost. A lost packet is typically detected and re-sent by a higher layer, so loss shows up for the user as even more delay. Loss and queuing delay are therefore two sides of one phenomenon: a link that cannot keep up.


Throughput and the bottleneck

So far we measured how long a packet takes. A different question is: how fast does data actually flow end to end? The answer is throughput, the number of bits per second that actually get from sender to receiver, measured at some moment (instantaneous throughput) or averaged over a long transfer (average throughput). If Mira downloads a file of (F) bits and the transfer takes (T) seconds, the average throughput is (F / T).

Throughput differs from link rate in an important way. A link's rate (R) is the best the link could do. Throughput is what the whole path delivers. Consider a path with links of rates (R_1, R_2, \dots, R_N) in series. Data cannot flow end to end faster than the slowest link allows, because the slow link limits how many bits per second get through. So, when only one flow uses the path and nothing else interferes, the throughput is

[ \text{throughput} = \min(R_1, R_2, \dots, R_N). ]

The slowest link is called the bottleneck link. Think of water flowing through pipes of different widths in a row: no matter how wide the others are, the flow out of the end is set by the narrowest.

When many flows share a link, the share each receives is smaller. If a core link of 2 Gbps carries 8 equal flows, each flow gets about 250 Mbps on that link, and that share is what you compare to the other links when finding the bottleneck. For most home users, the bottleneck is the access link, which is why upgrading the home connection often makes things faster; but if the access link becomes very fast, the bottleneck may move to a shared link elsewhere or to the server.

Common confusion: Throughput is not the same as the link's advertised rate, and it is not the same as the delay. A path can have very low delay (a short hop) and low throughput (a slow bottleneck), or high delay (a long satellite hop) and high throughput. Delay answers "how long for one packet?"; throughput answers "how many bits per second over time?"


Pulling the session together

Each hop of each packet contributes four delays: processing, queuing, transmission ((L/R)), and propagation ((d/s)). Transmission and propagation are fixed by link and packet; processing is tiny; queuing is the variable part, driven by traffic intensity (La/R). With store-and-forward, a path of (N) links pays (N) transmission delays and (N) propagation delays for a single packet. Cutting messages into packets lets links work in a pipeline. Throughput is the long-run rate of delivery, set by the slowest (or most heavily shared) link on the path.

In summary:

  • Four delays: processing, queuing, transmission (L/R), propagation (d/s).
  • Transmission depends on packet length and rate; propagation depends on distance and medium.
  • Queuing is the variable one; it explodes as traffic intensity approaches 1.
  • End-to-end delay sums the four delays over every hop, with one (L/R) per link.
  • Throughput of a path equals its bottleneck: the minimum rate (or smallest share).
  • Delay (time for one packet) and throughput (bits per second) are different measures.

Next session introduces the layers that organise all of this activity into manageable pieces.

Examples

Worked examples — delay, queues, and throughput on Mira's forecast path

Five calculations follow. For each, the Scene sets the stage, Numbers lists the inputs, and Working shows every line. Try the arithmetic yourself before reading on, and write the unit next to each number as you go.

Reminders: 1 byte = 8 bits; 1 Mbps = 1,000,000 bits per second; 1 ms = 0.001 s; propagation speed (s = 2 \times 10^8) m/s unless stated.


Example 1 — One link: transmission versus propagation

Scene. Mira's app sends a small request packet from her phone's nearby BayLink exchange to a SpineNet router on one long link. We want to know how long it takes for the whole packet to get across that single link, and what happens if the link is upgraded.

Numbers.

  • Packet length: 900 bytes.
  • Link rate: (R = 6) Mbps.
  • Link length: 1,800 km.
  • Propagation speed: (2 \times 10^8) m/s.

Working.

Step 1: convert to base units. (L = 900 \times 8 = 7{,}200) bits. (R = 6 \times 10^6) bits/s. (d = 1{,}800{,}000) m.

Step 2: transmission delay.

[ d_{\text{trans}} = \frac{7{,}200}{6{,}000{,}000} = 0.0012\ \text{s} = 1.2\ \text{ms}. ]

Step 3: propagation delay.

[ d_{\text{prop}} = \frac{1{,}800{,}000}{200{,}000{,}000} = 0.009\ \text{s} = 9\ \text{ms}. ]

Step 4: when does each moment happen? The sender starts at (t = 0). The first bit reaches the far end at (t = 9) ms, since a single bit only needs the propagation time. The last bit leaves the sender at (t = 1.2) ms, and it needs another 9 ms to cross, so it arrives at (t = 1.2 + 9 = 10.2) ms. The whole packet is received at 10.2 ms.

Step 5: upgrade the link to 600 Mbps. New transmission delay is (7{,}200 / 600{,}000{,}000 = 0.000012) s = 0.012 ms. Propagation is unchanged at 9 ms. The new arrival time of the last bit is (0.012 + 9 = 9.012) ms.

Interpretation. A hundredfold faster link saved only about 1.19 ms out of 10.2 ms, because on this long link almost all of the time is propagation. A faster link cannot shorten a long distance.

Takeaway. Compute both numbers separately and compare them. On short, slow links, transmission dominates; on long, fast links, propagation dominates.


Example 2 — Three links, store-and-forward, with and without a queue

Scene. A packet travels from Mira's phone to the Northline Relay delivery point through two routers: phone → R1 (BayLink) → R2 (SpineNet) → Northline. We compute the end-to-end delay when everything is quiet, and then when a short queue forms at R2.

Numbers.

| Link | Rate | Length | |------|------|--------| | 1 (phone → R1) | 20 Mbps | 3 km | | 2 (R1 → R2) | 8 Mbps | 640 km | | 3 (R2 → Northline) | 50 Mbps | 150 km |

Packet length: 3,000 bytes. Processing delay: 0.04 ms at each router. Queuing delay: 0 for the first case.

Working, first case (no queues).

Step 1: packet length in bits. (L = 3{,}000 \times 8 = 24{,}000) bits.

Step 2: transmission delay per link.

  • Link 1: (24{,}000 / 20{,}000{,}000 = 0.0012) s = 1.2 ms.
  • Link 2: (24{,}000 / 8{,}000{,}000 = 0.003) s = 3 ms.
  • Link 3: (24{,}000 / 50{,}000{,}000 = 0.00048) s = 0.48 ms.
  • Sum: (1.2 + 3 + 0.48 = 4.68) ms.

Step 3: propagation delay per link.

  • Link 1: (3{,}000 / 2 \times 10^8 = 1.5 \times 10^{-5}) s = 0.015 ms.
  • Link 2: (640{,}000 / 2 \times 10^8 = 0.0032) s = 3.2 ms.
  • Link 3: (150{,}000 / 2 \times 10^8 = 0.00075) s = 0.75 ms.
  • Sum: (0.015 + 3.2 + 0.75 = 3.965) ms.

Step 4: processing. There are two routers, each adding 0.04 ms: (2 \times 0.04 = 0.08) ms.

Step 5: total. (4.68 + 3.965 + 0.08 = 8.725) ms.

Working, second case (a queue at R2). Now suppose that when Mira's packet arrives at R2, three other packets of 24,000 bits each are already sitting in R2's buffer, waiting for link 3 (50 Mbps). Each of them takes 0.48 ms to send, so Mira's packet waits (3 \times 0.48 = 1.44) ms at R2 before its own transmission begins.

New total: (8.725 + 1.44 = 10.165) ms.

Interpretation. Link 2 is the slowest link (8 Mbps), and it contributes the largest single transmission delay, 3 ms, plus the largest propagation delay, 3.2 ms, because it is also the longest. Queueing added about 16% to the total here. In a heavily loaded network, the queuing term could easily be larger than the entire quiet-case total.

Takeaway. Write a small table of per-link transmission and propagation delays, sum each column, add processing and queuing, and the total follows. Each link contributes one of each; each router contributes processing and (possibly) queuing.


Example 3 — Cutting a message into packets

Scene. Mira's app uploads a small diagnostic log to the Skylark Forecast server. The path has two links in series with two equal rates, so one router sits between. We compare sending the log as one big packet versus five smaller ones.

Numbers.

  • Message: 240,000 bits (30,000 bytes).
  • Each link: 4 Mbps. Ignore propagation, processing, and queuing.

Working, one big packet.

Step 1: Transmission time per link = (240{,}000 / 4{,}000{,}000 = 0.06) s = 60 ms.

Step 2: With store-and-forward, link 2 cannot start until link 1 is completely done. Total = (60 + 60 = 120) ms.

Working, five equal packets.

Step 3: Each packet has (240{,}000 / 5 = 48{,}000) bits. Per-packet transmission time on one link = (48{,}000 / 4{,}000{,}000 = 0.012) s = 12 ms.

Step 4: timeline. Let (P_1 \dots P_5) be the packets.

Link 1 finishes P1 at 12 ms, P2 at 24 ms, P3 at 36 ms, P4 at 48 ms, P5 at 60 ms.
Link 2 starts P1 at 12 ms and finishes at 24 ms.
Link 2 can start P2 at 24 ms (P2 has arrived, and link 2 is free) -> finishes at 36 ms.
Link 2 finishes P3 at 48 ms, P4 at 60 ms, P5 at 72 ms.

Step 5: total. The last packet is fully delivered at 72 ms. As a check, ((P + K - 1) \times 12) with (P = 5) packets and (K = 2) links is (6 \times 12 = 72) ms. ✓

Interpretation. Splitting into five packets cut the time from 120 ms to 72 ms, a saving of 48 ms (40%). The second link started working at 12 ms instead of waiting until 60 ms.

Takeaway. More, smaller packets let links work in an overlapping pipeline. The gain is larger when there are more links, though real packets also carry headers that add overhead, so there are limits to how small packets should be.


Example 4 — Reading traffic intensity

Scene. A BayLink router's outgoing link serves a neighbourhood. The network team measures arrivals at busy times and wants to know if queues will stay manageable.

Numbers.

  • Link rate: (R = 8) Mbps.
  • Average packet length: (L = 5{,}000) bits.
  • Case A: arrival rate (a = 1{,}200) packets/s.
  • Case B: arrival rate (a = 1{,}800) packets/s.

Working.

Case A. Average bits arriving per second: (L \cdot a = 5{,}000 \times 1{,}200 = 6{,}000{,}000) bits/s. Traffic intensity = (6{,}000{,}000 / 8{,}000{,}000 = 0.75).

Case B. (L \cdot a = 5{,}000 \times 1{,}800 = 9{,}000{,}000) bits/s. Traffic intensity = (9{,}000{,}000 / 8{,}000{,}000 = 1.125).

Service rate in packets per second. The link sends (8{,}000{,}000 / 5{,}000 = 1{,}600) packets/s.

Case B growth. Arrivals are 1,800 packets/s and departures at most 1,600 packets/s, so the queue grows by (1{,}800 - 1{,}600 = 200) packets every second. If the buffer holds 400 waiting packets, it fills in (400 / 200 = 2) seconds, after which packets are dropped.

Interpretation. Case A (0.75) is loaded but stable: the link keeps up on average, though bursts will create short queues, and queuing delay will be noticeable. Case B (1.125) exceeds capacity, so the queue must grow until loss begins. A real operator would act on case A well before it creeps up toward 1.

Takeaway. Compute intensity as bits arriving per second over link rate. Below 1, queues are finite but grow quickly as the ratio nears 1; above 1, the buffer fills and packets are lost.


Example 5 — Finding the bottleneck

Scene. Mira downloads a map pack from edge.northline.example. The data crosses three links in series. Her friend asks how long it will take and whether paying for a faster home plan would help.

Numbers.

  • Link 1 (Northline server to SpineNet): 400 Mbps, used only by this transfer.
  • Link 2 (a shared core link inside SpineNet): 2,000 Mbps shared equally among 8 flows.
  • Link 3 (BayLink access to Mira's flat): 35 Mbps, used only by this transfer.
  • File: 70 megabytes.

Working.

Step 1: file size in bits. (70 \times 8 = 560) megabits.

Step 2: rate available on each link. Link 1: 400 Mbps. Link 2: (2{,}000 / 8 = 250) Mbps per flow. Link 3: 35 Mbps.

Step 3: bottleneck. (\min(400, 250, 35) = 35) Mbps. The access link is the bottleneck.

Step 4: download time. (560 / 35 = 16) s.

Step 5: try the faster home plan. Suppose Mira upgrades link 3 to 300 Mbps. The rates become 400, 250, and 300, and the minimum is now 250 Mbps (the core share). New time: (560 / 250 = 2.24) s.

Interpretation. The first upgrade is dramatic: from 16 s to 2.24 s. Notice the upgrade rate was 300 Mbps, but the result is capped at 250 Mbps because the bottleneck moved to the shared core link. Paying for an even faster plan, say 1,000 Mbps, would not help further: the time would stay at 2.24 s until the core share changed.

Takeaway. Throughput is the minimum rate along the path (counting shared links at their per-flow share). To speed a transfer up, find the bottleneck and improve that; improving a non-bottleneck link achieves nothing.

Practice check — delay, loss, and throughput

Multiple-choice check from the Matheion app lesson. Pick an answer for each question, then submit.

  1. 1.A packet of 2,500 bytes is placed on a 5 Mbps link. What is its transmission delay?

  2. 2.A fiber link is 1,600 km long and the signal travels at 2 x 10^8 m/s. What is the propagation delay?

  3. 3.The operator doubles a link's rate from 10 Mbps to 20 Mbps without changing its length. What happens to the delays for a given packet?

  4. 4.Mira measures the round-trip time of repeated identical requests along one fixed route: 61 ms, 62 ms, 118 ms, 61 ms, 143 ms, 62 ms. Which delay component best explains the large differences between measurements?

  5. 5.A packet of 18,000 bits crosses two store-and-forward links in series, the first at 3 Mbps and the second at 9 Mbps. Ignoring propagation, processing, and queuing, how long until the whole packet has been received at the end?

  6. 6.A single flow crosses links of 90 Mbps, 12 Mbps, and 40 Mbps in series, with nothing else using them. What is its throughput, and which link limits it?

  7. 7.A link sends at 3 Mbps. Packets average 8,000 bits and arrive at an average of 250 packets per second. What is the traffic intensity and what does it suggest?

  8. 8.Which statement correctly contrasts delay and throughput?