Computer Networking › Section A · Computer networks and the internet › Topic A.4
Delay, loss, and throughput
A network would ideally move any amount of data instantly and never lose a bit. Real networks delay packets, sometimes drop them, and deliver data only as fast as their slowest link allows. Three simple formulas explain most of what you notice as speed.
Four delays at every hop
As a packet travels from one node, a host or router, to the next, it picks up four kinds of delay. Processing delay is the time to examine the header, check for bit errors and choose the outgoing link, usually microseconds. Queueing delay is the time spent waiting in the output buffer behind other packets; it ranges from zero to many milliseconds. Transmission delay is the time to push all the packet's bits onto the link. Propagation delay is the time a bit takes to travel the length of the link. Their sum is the nodal delay:
Transmission versus propagation
These two are easy to confuse. Transmission delay is \(L/R\): packet length \(L\) bits divided by link rate \(R\) bit/s. It depends on the packet and the link's rate, not on the link's length. Propagation delay is \(d/s\): link length \(d\) metres divided by signal speed \(s\), which is about \(2 \times 10^8\) m/s in fibre and copper, roughly two-thirds of the speed of light in a vacuum. It depends on distance, not on the packet or the rate. Think of it this way: transmission is how long the last bit waits to leave; propagation is how long each bit spends in transit.
Worked example: a short fast link and a long one
Compare a 1,500-byte packet on two links, with \(s = 2 \times 10^8\) m/s. 1. Packet size: \(L = 1{,}500 \times 8 = 12{,}000\) bits. 2. A 100 m office link at 10 Mbps: transmission \(12{,}000 \div 10^7 = 1.2\) ms; propagation \(100 \div (2 \times 10^8) = 0.5\) µs. 3. A 3,000 km fibre at 10 Gbps: transmission \(12{,}000 \div 10^{10} = 1.2\) µs; propagation \(3 \times 10^6 \div (2 \times 10^8) = 15\) ms. 4. On the short link transmission dominates; on the long one propagation dominates. Check: making the long link 10 times faster would cut only the 1.2 µs, leaving the 15 ms untouched.
Traffic intensity and queueing
Queueing delay changes from packet to packet, so it is described statistically. If packets of \(L\) bits arrive at a queue at an average rate of \(a\) packets per second and leave on a link of rate \(R\), the traffic intensity is \(La/R\). If \(La/R > 1\), bits arrive faster than the link can send them, and the queue grows without limit; a well-designed system keeps it below 1. Even below 1, bursts build temporary queues, and the average delay rises steeply as the intensity approaches 1. For example, 700 packets/s of 12,000 bits on a 10 Mbps link give \(8.4 \times 10^6 \div 10^7 = 0.84\). Oversized buffers that let queues grow huge cause bufferbloat: long delays without loss.
Packet loss
A buffer has finite capacity. A packet that arrives at a full buffer has nowhere to wait, so the router drops it, and the packet is lost. The fraction of packets lost rises as traffic intensity grows, so performance is judged by both delay and loss probability. A lost packet may be resent by the source, as TCP does to guarantee delivery, or simply skipped, as many live voice and video applications prefer, since a late packet is useless to them. Rising delay together with loss is the classic sign of congestion, studied in the topic Principles of congestion control.
End-to-end delay, RTT and measuring tools
If a path has \(N\) links, each with the same delays and no queueing, the end-to-end delay is \(N(d_{\text{proc}} + d_{\text{trans}} + d_{\text{prop}})\); on real paths, add each hop's delays separately. The round-trip time (RTT) is the time for a small packet to reach a destination and for the reply to return. The ping tool reports RTTs; traceroute sends probes with increasing time-to-live values, so each router in turn reports back, revealing the hops and the RTT to each. Variation in delay from packet to packet is called jitter, which matters for calls and games. Hosts can add delays of their own, for example while waiting for a shared channel.
Worked example: delay across three hops
A 1,250-byte packet goes host, router R1, router R2, server. Links: 100 Mbps over 50 m; 1 Gbps over 800 km; 1 Gbps over 200 km. Each router takes 20 µs to process; assume empty queues and \(s = 2 \times 10^8\) m/s. 1. \(L = 10{,}000\) bits. 2. Transmission: \(100 + 10 + 10 = 120\) µs. 3. Propagation: \(0.25\) µs \(+ 4\) ms \(+ 1\) ms \(= 5.00025\) ms. 4. Processing: \(2 \times 20 = 40\) µs. 5. Total: \(0.12 + 5.00025 + 0.04 \approx 5.16\) ms. Check: propagation is about 97% of the total, so a faster link would barely help; a shorter path would.
Throughput and the bottleneck link
Throughput is the rate, in bits per second, at which the receiver actually gets data. Instantaneous throughput is the rate at one moment; average throughput is the file size divided by the time taken. With no other traffic, a path of links with rates \(R_1, R_2, \ldots, R_N\) delivers
The slowest link is the bottleneck. A link shared by \(n\) busy flows gives each roughly \(R/n\). Today the bottleneck is usually the access network, but a crowded shared link can take that role. The bandwidth-delay product, rate times RTT, is the number of bits in flight on a path: 100 Mbps with a 40 ms RTT holds \(4 \times 10^6\) bits, or 0.5 MB.
Worked example: finding the bottleneck
A server with a 1 Gbps link sends a 2 GB file to a client whose access link is 300 Mbps. In between, both cross a 2 Gbps core link that is shared equally by 20 busy flows. 1. Server link: 1,000 Mbps. 2. Core share: \(2{,}000 \div 20 = 100\) Mbps. 3. Client link: 300 Mbps. 4. Throughput: \(\min(1{,}000, 100, 300) = 100\) Mbps, so the shared core link is the bottleneck. 5. Time: \(\frac{2 \times 10^9 \times 8}{10^8} = 160\) s. Check: upgrading the client to 1 Gbps would not help at all; only a lighter load or more core capacity would.
Common mistakes to avoid
First, swapping \(L/R\) and \(d/s\): a faster link shortens transmission, never propagation. Second, forgetting to convert bytes to bits before dividing by a rate. Third, using the speed of light in a vacuum for fibre instead of about \(2 \times 10^8\) m/s. Fourth, adding link rates or averaging them to find throughput; take the minimum. Fifth, thinking a traffic intensity just below 1 is safe; queueing delay is already very large there. Sixth, treating loss as random bad luck rather than a symptom of full buffers. Seventh, mixing up RTT, which is there and back, with one-way delay.
Key takeaways
Each hop adds processing, queueing, transmission and propagation delay. Transmission is \(L/R\) and depends on rate; propagation is \(d/s\), with \(s \approx 2 \times 10^8\) m/s in fibre, and depends on distance. Traffic intensity \(La/R\) must stay below 1, and delay soars as it nears 1; oversized buffers cause bufferbloat. Full buffers drop packets, which TCP resends and live media may skip. End-to-end delay sums each hop's delays; ping measures RTT, traceroute shows the hops, and jitter is delay variation. Throughput is set by the bottleneck, the minimum link rate along the path, with shared links giving each flow about \(R/n\). Rate times RTT gives the bits in flight.
Flashcards
Flashcard
Processing + queueing + transmission + propagation
Test your understanding
Question 1 of 5
What is the propagation delay over 4,000 km of fibre, where signals travel at about 2 × 10⁸ m/s?
Question 2 of 5
A 1,000-byte packet is sent on a 100 Mbps link that is 20 km long. What is its transmission delay?
Question 3 of 5
Packets of 10,000 bits arrive at a router queue at an average of 1,200 per second, and the outgoing link runs at 10 Mbps. What happens to the queue?
Question 4 of 5
A file travels over links of 500 Mbps, 50 Mbps and 1 Gbps, with no other traffic. What is the end-to-end throughput?
Question 5 of 5
A user's video downloads slowly, and the path's bottleneck is a 1 Gbps core link shared equally by 25 busy flows. The user's own access link is 200 Mbps. What would raise this user's throughput?
