Computer Networking › Section G · Wireless and mobile networks › Topic G.7
Wireless and mobility: impact on higher layers
TCP, UDP and applications run unchanged over Wi-Fi, cellular and satellite links, yet they can perform very differently there. The reason is that wireless losses and delays send misleading signals to protocols designed for wires.
Same protocols, different behaviour
The network layer offers the same best-effort service over wireless links as over cables, so TCP and UDP work without modification, and so do applications. In that sense the layering does its job. The differences appear in performance. Wireless links lose and corrupt more packets, their delay swings as signal quality and scheduling change, handovers can drop or delay packets in flight, and some links, such as satellites, add long delays. Protocols above the link layer react to these effects, sometimes badly. This topic looks at why, and at the fixes used in practice at the link, transport and application layers.
TCP reads every loss as congestion
As "TCP congestion control" explains, a TCP sender treats a lost segment as a sign that a router queue overflowed, and it cuts its congestion window, lowering its sending rate. On a wired path that is usually right. On a wireless path, a segment may instead be lost to bit errors or during a handover while the routers' queues are nearly empty. TCP cannot tell the difference: a missing acknowledgement looks the same whatever the cause. So it slows down when it should simply resend, and the link sits underused. The more often wireless losses occur, the further TCP's rate falls below what the path could carry.
Worked example: how loss limits TCP
A widely used approximation for a long-running TCP flow with segment loss probability \(p\) is
It follows from the Reno sawtooth in "TCP congestion control". Take MSS = 1,460 bytes and RTT = 50 ms. 1. MSS in bits: 1,460 × 8 = 11,680. 2. Wired path, \(p = 10^{-4}\): 1.22 × 11,680 ÷ (0.05 × 0.01) ≈ 28.5 Mbit/s. 3. Lossy wireless path, \(p = 10^{-2}\): the square root becomes 0.1, so the rate is about 2.85 Mbit/s. 4. Interpret: a hundredfold rise in loss cuts throughput tenfold, even if no router is congested. Check: the rate depends on \(1/\sqrt{p}\), so quadrupling the loss halves it.
Fix one: recover locally
The most common remedy hides wireless losses from TCP by repairing them on the link where they happen. 802.11 retransmits any frame that is not acknowledged. LTE and 5G combine hybrid ARQ, which adds error-correcting redundancy and retransmits within milliseconds, with RLC retransmission above it. TCP is never told a wireless link exists; it simply sees far fewer losses. The cost is extra and variable delay, because a frame that needs several attempts arrives later than its neighbours, and retransmissions use airtime that new data could have used.
Worked example: retries hide loss
A radio link loses 10% of frames on each attempt, independently, and the link layer allows up to 3 retransmissions, 4 attempts in all. What loss does TCP see? 1. A frame is lost only if every attempt fails. 2. Probability: \(0.1^4 = 10^{-4}\), or 0.01%. 3. Compare the previous example: a loss rate of \(10^{-4}\) supported about 28.5 Mbit/s, while 10% loss seen directly by TCP would be crippling. 4. Average attempts per frame: about 1.11, so repeats add roughly 11% to the airtime used. Check: real losses are often bursty rather than independent, so retries help a little less than this ideal.
Fix two: smarter transport
A second approach makes the transport layer aware that not every loss means congestion. Explicit Congestion Notification lets routers mark packets when queues build up, so a sender can learn of congestion without relying on loss. Congestion-control algorithms such as BBR, from Google and covered in "Evolution of the transport layer: QUIC", estimate the path's bottleneck rate and round-trip time and react much less to isolated losses than classic loss-based TCP. Research designs have also tried to tell wireless losses from congestion losses directly. These methods keep end-to-end semantics, but sender and receiver software must be updated.
Fix three: split the connection
In a split-connection design, a proxy at the edge of the wireless network ends the mobile device's TCP connection and opens a second one to the server. Each half is tuned to its own path: the wired half sees almost no loss, while the wireless half can use local recovery or a tailored protocol. Such performance-enhancing proxies have been widely used by cellular and satellite operators. They have drawbacks: the proxy must keep state for every connection, a segment acknowledged by the proxy may never reach the server, and they cannot split encrypted transports such as QUIC, whose packet numbers and acknowledgements are encrypted.
Moving devices and changing addresses
When a device switches from Wi-Fi to cellular, its IP address changes, and a TCP connection, tied to the two addresses and ports, breaks. Two transport-level answers avoid that. QUIC identifies each connection by a connection ID rather than by addresses, so a connection can migrate to a new address and carry on, as described in "Evolution of the transport layer: QUIC". Multipath TCP, specified in RFC 8684, lets one connection use several paths at once, such as Wi-Fi and cellular, and keep going if one disappears. Within one cellular network, tunnels already keep the address fixed, as "Mobility management in practice" showed.
Worked example: TCP over a satellite
A geostationary satellite orbits 35,786 km above the equator. A request travels ground to satellite to ground, and the reply does the same: four legs. 1. Minimum RTT: \(4 \times 35{,}786{,}000 \div (3 \times 10^8) \approx 0.477\) s. 2. Without window scaling, TCP can have at most 65,535 bytes unacknowledged, so the rate is at most 65,535 × 8 ÷ 0.477 ≈ 1.1 Mbit/s, however fast the link. 3. Filling a 50 Mbit/s link needs a window of 50 × 10⁶ × 0.477 ÷ 8 ≈ 3.0 MB, which window scaling allows. 4. A low Earth orbit satellite at roughly 550 km gives a minimum RTT of about 7.3 ms, but a constellation needs thousands of satellites and frequent handovers between them. Check: real RTTs are higher once ground networks are added.
Effects on applications
Applications feel wireless links too. Bandwidth is shared and changes as users move, so video players use adaptive bitrate streaming, switching between quality levels as throughput rises and falls, as "Video streaming and content distribution networks" describes. Well-designed mobile apps send compact data, batch requests so the radio can sleep longer, and keep working offline, synchronising when a connection returns. Battery life matters as much as speed, because waking the radio for every small message is costly. Mobility also opens possibilities: location-aware services such as navigation, ride hailing and local search exist because the device moves with its user.
Common mistakes to avoid
First, assuming that because TCP works over wireless it works well; it often underuses the link. Second, blaming every wireless slowdown on congestion when bit errors and handovers cause loss too. Third, thinking more link-layer retries are free; they add delay variation and consume airtime. Fourth, treating a fast satellite link as low latency: a geostationary RTT is at least about 0.48 s whatever the bit rate. Fifth, forgetting that split-connection proxies break end-to-end semantics and cannot see inside QUIC. Sixth, mixing bits and bytes in window calculations.
Key takeaways
TCP and UDP run unchanged over wireless, but performance differs. TCP treats every loss as congestion and slows down, so wireless bit errors and handover losses cut throughput, roughly as 1/√p. Fixes work at three levels: local link-layer recovery (802.11 ARQ, hybrid ARQ, RLC), congestion signals or algorithms less tied to loss (ECN, BBR), and split connections through proxies. QUIC connection IDs and Multipath TCP let connections survive address changes. A geostationary hop adds at least about 477 ms of RTT, so windows must scale. Applications adapt bitrates, save battery and use location.
Flashcards
Flashcard
TCP works unchanged but can perform poorly
Test your understanding
Question 1 of 5
A phone's Wi-Fi link loses some segments to bit errors, though no router on the path is congested. Why does its TCP throughput still fall?
Question 2 of 5
Using rate ≈ 1.22 × MSS ÷ (RTT × √p), what happens to a TCP flow's rate if the loss probability p becomes four times larger, with RTT unchanged?
Question 3 of 5
A link loses 20% of frames per attempt, independently, and allows up to 2 retransmissions (3 attempts in all). What fraction of frames does TCP see lost?
Question 4 of 5
A satellite path has an RTT of 0.6 s, and the TCP window is 1,000,000 bytes. What is the most the connection can carry?
Question 5 of 5
A phone walks out of Wi-Fi range onto cellular, gaining a new IP address. Why can a QUIC connection carry on when a plain TCP one breaks?
