Course: CS176C — Advanced Topics in Internet Computing, Spring 2026
Instructor: Arpit Gupta, UC Santa Barbara
Date: April 14, 2026
Slides: Deployed slide deck
Pre-requisite: L4 (SDN and the four-invariant framework applied to familiar systems)
A new domain where the framework meets physics
Lectures 1 through 4 built the four-invariant framework — State, Time, Coordination, Interface — on systems you already knew: TCP, DHCP, DNS, routing, BGP, SDN. In every one of those systems, a sender could observe outcomes directly. A router sees its queue depth. A TCP sender receives ACKs. The feedback loop was always there, waiting to be analyzed.
Today that changes. We enter a domain where physics denies the most basic form of feedback: you literally cannot hear what is happening to your own transmission. The shared-medium problem is the oldest problem in networking — older than the Internet itself — and it forces protocol designers to confront a gap that no amount of software can close: the environment is global, but measurement is local. What you hear at your antenna tells you about your neighborhood, not the network. Every medium access failure lives in that gap.
The lecture traces the solution from nothing — Pure ALOHA in 1970 [1] — through carrier sensing in 1975 [3] to collision detection on Ethernet in 1976 [4]. Each step adds one piece of shared state to the system. Each achieves higher utilization. And each breaks under conditions that the next generation must address. By the end, the assumption cascade hits a physics wall: wireless cannot detect collisions. How wireless closes the feedback loop without collision detection is the question that Lecture 6 answers.
The four physical facts you cannot engineer away
A finite radio channel must carry transmissions from many devices. Transmission is broadcast — every station in range hears every signal. A transmitter cannot hear its own collision, because 40–50 dB of self-interference saturates the receiver. And carrier sensing is location-dependent: what you hear depends on where you are.
From these facts emerges the core gap. Suppose station A transmits while station C also transmits. From A’s perspective the channel appears idle — A hears only its own signal. But station B, positioned between A and C, receives both signals simultaneously. Both frames are destroyed at B. The environment is global. Measurement is local. This asymmetry is where every medium access failure lives.
Why radio behaves this way
Three physical facts and one institutional fact determine the design space for wireless medium access.
Path loss. The Friis equation gives free-space loss proportional to the square of both frequency and distance: $L \propto f^2 \cdot d^2$ [5]. Doubling distance costs 6 dB. Doubling frequency costs another 6 dB at the same distance. Indoors the situation is worse — walls, floors, and multipath reflections push decay toward $1/d^3$ or $1/d^4$.
The frequency–range–bandwidth tradeoff. Low frequencies propagate farther because they suffer less free-space loss and diffract around obstacles. High frequencies offer more capacity because available bandwidth scales with carrier frequency, and smaller antennas at high frequency enable beamforming and massive MIMO. The following table captures how real systems navigate this tradeoff:
| Band | Range | Bandwidth | Antenna | Used by |
|---|---|---|---|---|
| 900 MHz | km | Narrow | 16 cm | Early cellular |
| 2.4 GHz | ~50 m | 85 MHz | 6 cm | WiFi b/g/n |
| 5 GHz | ~30 m | 500 MHz | 3 cm | WiFi a/ac/ax |
| 6 GHz | ~25 m | 1.2 GHz | 2.5 cm | WiFi 6E |
| 28 GHz | ~200 m LOS | GHz+ | 5 mm | 5G mmWave |
SNR determines modulation. Received power equals transmit power minus path loss; the noise floor sits at approximately −90 dBm. A receiver needs SNR above 10 dB for BPSK (the simplest modulation) and above 25 dB for 256-QAM (the densest practical modulation). This is why WiFi “slows down” as you walk away from the access point — rate adaptation drops the modulation order as SNR falls.
Spectrum licensing drives protocol design. Unlicensed bands (2.4, 5, 6 GHz) are open to anyone — multiple transmitters contend for the channel, producing contention-based protocols like WiFi. Licensed bands (cellular frequencies) grant one operator exclusive access via government auction — the base station is the sole scheduler, and contention-based access is unnecessary. This institutional split is the fork in the road between WiFi and cellular, and it shapes everything that follows in Lectures 6 through 8.
Hawaii, 1969: the birth of contention-based access
Norman Abramson faced a problem with no existing solution. The University of Hawaii had campuses on four islands. One shared radio channel connected them. There were no wires between islands — radio was the only option. There was no central controller, no shared clock, and no way to detect a collision during transmission. Under these constraints, the simplest viable protocol was: transmit whenever you have data, deal with collisions after the fact.
Pure ALOHA and the vulnerable window
Pure ALOHA’s rule is minimal: transmit whenever you have data; if a collision occurs (detected by the absence of an acknowledgment), wait a random time and retry [1].
The critical insight is the vulnerable window. Your frame takes $T$ seconds to transmit. Any other station that began transmitting within the interval from $T$ seconds before your start to $T$ seconds after your start will overlap with your frame. The vulnerable window is therefore $2T$. With many independent, rare transmitters, aggregate arrivals follow a Poisson process, and the probability of success — zero other arrivals in the vulnerable window — yields the throughput formula:
\[S = G \cdot e^{-2G}, \quad \text{peak at } G=0.5: \textbf{18.4\%}\]where $G$ is the offered load in frames per frame-time [1]. Over 80% of the channel is wasted. With zero shared state, this is a hard ceiling.
The lasting contribution of ALOHA is not the equation but the concept of the vulnerable window. Every subsequent protocol in this lecture shrinks that window by adding one new dimension of shared state.
Slotted ALOHA: one bit of global state doubles throughput
Lawrence Roberts observed in 1972 that ALOHA’s vulnerable window could be halved by synchronizing all transmitters to a common slot boundary [2]. If frames can only begin at slot edges, partial overlaps become impossible. A collision means two frames in the same slot; otherwise the channel is clean. The vulnerable window drops from $2T$ to $T$, and the throughput formula becomes:
\[S = G \cdot e^{-G}, \quad \text{peak at } G=1.0: \textbf{36.8\%}\]The cost is one bit of global state — the slot boundary itself. Someone must broadcast a synchronization signal. But that single bit of shared state exactly doubled the peak throughput. The pattern is now visible: halving the vulnerable window doubles capacity. Each generation of medium access protocol adds one new dimension of shared state and achieves one corresponding improvement.
The instability of blind transmission
Slotted ALOHA had a deeper flaw beyond its 36.8% ceiling. Kleinrock and Lam observed two fatal problems [3]. First, instability under load: as traffic increases, collisions increase, which causes retransmissions, which cause more collisions. This positive feedback loop can collapse throughput to zero. The channel has a bistable equilibrium — it randomly flips between a functioning state and a collapsed state.
Second, stations ignore available information. On a local-area network, propagation delay $\tau$ is approximately 5 microseconds while frame time $T$ is approximately 1 millisecond. Station B can hear A’s transmission almost instantly — the information is free. But ALOHA never asks B to listen. Stations transmit into a channel they could have known was busy, if only they had checked.
Carrier sensing: exploiting free information
The fix was obvious once stated. Carrier Sense Multiple Access (CSMA), formalized by Kleinrock and Tobagi in 1975, adds one rule: listen before transmitting [3]. Carrier sensing means detecting energy on the medium — a signal above the noise floor indicates someone is transmitting. If the channel is busy, defer. If idle, transmit immediately (1-persistent CSMA) or with probability $p$ ($p$-persistent CSMA).
The vulnerable window shrinks from $T$ (the slot time in Slotted ALOHA) to $\tau$ (the propagation delay). On a LAN where $T \approx 1$ ms and $\tau \approx 5$ μs, this is a 200× reduction. Throughput jumps to 80–90%, compared to Slotted ALOHA’s 36.8% [3].
The new shared state is channel busy/idle — a single bit, but one acquired through local measurement rather than broadcast synchronization.
But the assumption underlying CSMA can fail. “If I hear idle, it is idle everywhere” is wrong. If A starts transmitting and B is one propagation delay away, B still senses idle because A’s signal has not yet arrived. B transmits — collision. The residual vulnerability is exactly $\tau$, the propagation delay [3]. On a LAN this is tiny. On a satellite link it would be catastrophic. The ratio $\tau / T$ determines whether carrier sensing is worth anything at all.
Collision detection on the wire
On a wire, signals combine at every point. A transmitter can listen to the medium while transmitting because its own signal power (~1 mW, or 0 dBm) and the incoming signal power are comparable — roughly a factor of 2. Metcalfe and Boggs exploited this in 1976 [4]: sense the channel; if idle, start transmitting; monitor the wire while sending; if the received signal mismatches what you sent, a collision has occurred — abort immediately. Send a 48-bit jam signal so all parties detect the collision, then enter Binary Exponential Backoff: after the $m$-th collision, pick $K \in {0,\ldots,2^m-1}$ and wait $K \times 512$ bit-times before retrying.
The key numbers at 10 Mbps tell the story:
| Parameter | Value |
|---|---|
| 512 bit-times | 51.2 μs |
| Collision detected and resolved in | microseconds |
| Bytes wasted per collision | A few |
| Utilization | ~95%+ |
CSMA/CD closes the feedback loop at the transmitter. The sender detects the collision on the wire and aborts within microseconds, wasting only a few bytes. Utilization reaches roughly 95% [4]. The assumption cascade has reached its zenith on guided media.
The physics wall: why wireless cannot detect collisions
Wireless breaks CSMA/CD for a reason rooted in the physics of unguided media, not in cost or design choice [5].
| Property | Guided medium (wire) | Unguided medium (radio) |
|---|---|---|
| Signal confinement | Confined to conductor | Radiates in all directions |
| TX power | ~1 mW (0 dBm) | ~100 mW (20 dBm) |
| Other stations’ signal | Comparable power | −60 dBm from 50 m |
| Own signal at antenna | 0 dBm | −10 dBm |
| Ratio (own vs. other) | ~2× | 100,000× (50 dB gap) |
| Collision detection | Works | Impossible — receiver saturated |
The causal chain is short: an unguided medium requires high transmit power to reach distant receivers; that high power at the transmitter’s own antenna drowns any incoming signal; the receiver cannot sense the medium while transmitting; collision detection is physically impossible; and so wireless must avoid collisions rather than detect them [5].
This is fundamental physics, not a tradeoff that future hardware might resolve. The 50 dB self-interference gap is inherent to radio transmission.
The assumption cascade in summary
Each generation of medium access protocol broke the previous generation’s binding assumption by adding one dimension of shared state:
| Generation | Shared State Added | Throughput | Key Assumption Broken |
|---|---|---|---|
| Pure ALOHA [1] | Nothing | 18.4% | “Transmit whenever” |
| Slotted ALOHA [2] | Slot boundary | 36.8% | “Align to slots” |
| CSMA [3] | Channel busy/idle | ~80–90% | “If idle here, idle everywhere” |
| CSMA/CD (wired) [4] | Collision signal | ~95%+ | “Can hear own collision” |
Wireless cannot achieve the last step. The assumption cascade hits a physics wall. The framework maps cleanly onto this evolution:
| Invariant | ALOHA [1] | CSMA [3] | CSMA/CD [4] |
|---|---|---|---|
| State | None | Channel busy/idle | Collision signal on wire |
| Time | Continuous (Pure) / Slotted | Propagation-delay limited | Microsecond detection |
| Coordination | Random retransmit | Listen-before-talk | Detect + abort + jam |
| Interface | Best-effort broadcast | Improved by sensing | Near-perfect with BEB |
When carrier sensing breaks: the τ/T ratio as architectural driver
CSMA works because $\tau \ll T$ — you hear others almost instantly relative to the frame time. What happens when that condition fails?
Geostationary satellite. Propagation delay to geostationary orbit and back is $\tau = 125$ ms. A typical frame takes $T \approx 5$ ms to transmit. The ratio $\tau / T = 25$. A station senses signals that are 25 frame-times old. By the time it hears “busy,” 25 frames could have started and finished. Carrier sensing is useless — the system is back to ALOHA. The natural solution is centralized scheduling: the satellite allocates time slots (TDMA), exactly as real systems like DVB-RCS implement.
Underwater acoustic network. Sound travels at 1,500 m/s. Sensors 1 km apart experience $\tau = 670$ ms with $T = 100$ ms, giving $\tau / T = 6.7$. Carrier sensing is again useless. But there is no infrastructure underwater — no satellite, no base station, no coordinator. With no central authority and useless sensing, these networks are stuck near ALOHA performance. This remains an open research problem.
The structural insight is that the choice between contention and scheduling is driven by physics, not ideology [3][5]:
| Regime | Sensing | Architecture |
|---|---|---|
| $\tau \ll T$ (LAN) | Works | Contention protocols (CSMA) |
| $\tau \gg T$ + central authority (satellite) | Useless | Centralized scheduling (TDMA) |
| $\tau \gg T$ + no authority (underwater) | Useless | Stuck near ALOHA (open problem) |
This connects backward: CSMA works on LANs because $\tau / T \approx 0.005$. It connects forward: WiFi density eventually breaks contention even when $\tau / T$ is small, forcing centralization for a different reason — collision probability under density, not sensing staleness. Lecture 6 takes up that story, beginning with how CSMA/CA closes the feedback loop at the receiver when collision detection is impossible, and tracing the consequences of that 12 ms frame-time-bounded loop through the throughput ceiling that ultimately demanded centralization.
References
[1] N. Abramson, “The ALOHA System — Another Alternative for Computer Communications,” Proc. Fall Joint Computer Conference, AFIPS, pp. 281–285, 1970.
[2] L. G. Roberts, “ALOHA Packet System with and without Slots and Capture,” ACM SIGCOMM Computer Communication Review, vol. 5, no. 2, pp. 28–42, April 1975. (Original slotted ALOHA concept ~1972.)
[3] L. Kleinrock and F. A. Tobagi, “Packet Switching in Radio Channels: Part I — Carrier Sense Multiple-Access Modes and Their Throughput-Delay Characteristics,” IEEE Trans. Communications, vol. COM-23, no. 12, pp. 1400–1416, Dec. 1975.
[4] R. M. Metcalfe and D. R. Boggs, “Ethernet: Distributed Packet Switching for Local Computer Networks,” Communications of the ACM, vol. 19, no. 7, pp. 395–404, July 1976.
[5] F. A. Tobagi and L. Kleinrock, “Packet Switching in Radio Channels: Part II — The Hidden Terminal Problem in Carrier Sense Multiple-Access and the Busy-Tone Solution,” IEEE Trans. Communications, vol. COM-23, no. 12, pp. 1417–1433, Dec. 1975.