The Shared-Medium Problem

CS176C — Advanced Topics in Internet Computing

Arpit Gupta

2026-04-14

Where We Are and Where We’re Going

Lectures 1–4: You built the four-invariant framework on systems you already knew — TCP, DHCP, DNS, routing, BGP, SDN.

Today: A new domain where the framework meets physics.

Sharing a medium is the oldest problem in networking — older than the Internet
We trace the solution from nothing (ALOHA, 1970) through carrier sensing (CSMA, 1975) to collision detection (Ethernet, 1976)
Each step adds one piece of shared state — and each breaks under new conditions

By the end of this lecture: You will understand the assumption cascade that drives medium access evolution — and why it hits a physics wall when you move from wires to radio.

Part 1: Foundations

ALOHA to CSMA/CA: The Assumption Cascade

The Shared-Medium Problem

The physics 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 (40–50 dB self-interference)
Carrier sensing is location-dependent — what you hear depends on where you are

    A's view         B's view         Reality
   ┌────────┐      ┌────────┐      ┌────────┐
   │ "idle" │      │ A + C  │      │ A + C  │
   │        │      │COLLIDE │      │COLLIDE │
   └────────┘      └────────┘      └────────┘
   A transmits     B receives       Both frames
   can't hear C    both signals     destroyed

        Environment is global.
        Measurement is local.

This gap between environment and measurement is where every medium access failure lives.

Why Radio Behaves This Way

Path loss — signal decays with distance AND frequency:

Friis equation: \(L \propto f^2 \cdot d^2\)
Doubling distance = 6 dB more loss
Doubling frequency = 6 dB more loss (at same distance)
Indoors worse: walls, floors, multipath add \(1/d^3\) to \(1/d^4\)

Why low frequency propagates farther:

Less free-space loss (lower \(f\) in \(f^2 d^2\))
Low-freq waves diffract around obstacles (wavelength \(\approx\) object size)
High-freq waves are blocked (wavelength \(\ll\) object size)
Concrete: a hand blocks 28 GHz; a wall blocks 5 GHz; 900 MHz passes through both

Why high frequency offers more capacity:

Bandwidth \(\propto\) carrier: 10% of 900 MHz = 90 MHz; 10% of 28 GHz = 2.8 GHz
More unallocated spectrum at higher bands
Smaller antenna (\(\propto\) wavelength) → more elements in same space → beamforming, massive MIMO

The tradeoff in practice:

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 what you can do:

Received power = TX power \(-\) path loss
Noise floor \(\approx -90\) dBm
Need SNR \(>\) 10 dB (BPSK) to \(>\) 25 dB (256-QAM)
WiFi “slows down” as you walk away: rate adaptation drops modulation as SNR falls

Spectrum licensing drives protocol design:

Unlicensed (2.4/5/6 GHz): anyone transmits → contention (WiFi)
Licensed (cellular bands): exclusive access → scheduling

Spectrum availability drives design:

2.4 GHz: unlicensed → anyone can use → must handle interference → contention protocols (WiFi)
Cellular bands: licensed → exclusive access → scheduling protocols :::

:::

This slide grounds the “shared-medium problem” in RF physics before jumping to protocols. Three concepts students need:

Path loss: electromagnetic signals weaken as they spread. Free-space inverse-square law gives 20 dB loss per 10x distance. Indoors it is worse (30-40 dB per decade) due to walls, floors, and multipath reflections. This is WHY carrier sensing is location-dependent: if station C is far enough from A, A’s signal is below C’s noise floor and C literally cannot detect it.
Frequency-range-bandwidth tradeoff: lower frequencies propagate farther and penetrate better (900 MHz cellular reaches km; 28 GHz mmWave is blocked by a hand). But spectrum below 1 GHz is extremely scarce — only a few tens of MHz available. Higher bands offer GHz of bandwidth but require line-of-sight. Antenna size is proportional to wavelength: a half-wave dipole at 2.4 GHz is ~6 cm (fits in a phone), at 900 MHz it is ~16 cm (still ok), at 60 GHz it is ~2.5 mm (tiny — enables massive arrays). At 28 GHz, phones use phased arrays of 4-8 elements to form beams.
SNR determines what you can do: the receiver needs the signal above the noise floor by enough margin to decode. Higher modulation (64-QAM, 256-QAM) requires higher SNR (>25 dB). Lower SNR forces simpler modulation (BPSK) at lower data rates. This is why WiFi “slows down” as you walk away from the AP — the rate adaptation algorithm drops to lower modulation as SNR decreases.
Spectrum licensing is the institutional anchor for the WiFi vs. cellular split: 2.4/5/6 GHz bands are unlicensed (ISM bands), so anyone can transmit — forcing contention-based protocols. Cellular bands are licensed via multi-billion-dollar auctions, giving operators exclusive access — enabling scheduling-based protocols. This distinction drives the entire narrative of L6.

Keep this slide to 3-4 minutes max. It is grounding, not the main content.

Hawaii, 1969: You Are Norman Abramson

The setup:

University of Hawaii has campuses on four islands
One shared radio channel connects them all
No wires between islands — radio is the only option
You need to let any campus send data to any other

The constraints:

No central controller (who would run it?)
No way to “listen” for collisions mid-transmission
Signals propagate at the speed of light — no time to react

graph TD
    A["Oahu"] ---|radio| S["Shared Channel"]
    B["Maui"] ---|radio| S
    C["Big Island"] ---|radio| S
    D["Kauai"] ---|radio| S
    style S fill:#e74c3c,stroke:#c0392b,color:white

Turn to your neighbor: design a protocol. How do you decide when to transmit?

Pure ALOHA (Abramson, 1970)

What did your group propose? Most of you said some version of “just send it.”

That’s exactly what Abramson did.

Protocol: Transmit whenever you have data. If collision, wait random time, retry.

Time ──────────────────────────────────►

       t-T        t        t+T
        │          │▓▓▓▓▓▓▓▓│
        │          │ YOUR   │
        │          │ FRAME  │
        │◄──────── 2T ─────►│
        │ VULNERABLE WINDOW │
        │                   │
  ▲ overlap            ▲ overlap       ▲ safe
  (they finish late)   (they start early)

The throughput formula:

\(G\) = offered load (frames per frame-time)
Many independent, rare transmitters → aggregate arrivals are Poisson (like raindrops on a window)
Success = zero other arrivals in vulnerable window
General form: \(S = G \cdot e^{-G \cdot (\text{window}/T)}\)

For Pure ALOHA: window = \(2T\)

\[S = G \cdot e^{-2G}, \quad \text{peak at } G=0.5: \textbf{18.4\%}\]

Over 80% wasted. With zero shared state, this is a hard ceiling.

Build the intuition step by step before showing any equation:

“Suppose your frame takes T seconds. For your frame to succeed, NO other station can be transmitting during your frame. But it’s worse than that — if another station started transmitting T seconds BEFORE you, their frame is still in the air when yours starts. So the danger zone starts at t-T.”
“Similarly, if another station starts transmitting at any point during your frame (from t to t+T), you collide. So the danger zone is [t-T, t+T] — a window of 2T.”
“Now the math: if other stations generate frames as a Poisson process with total rate G frames per frame-time, the probability of zero arrivals in a window of 2T is e^{-2G}. Throughput S = G × e^{-2G} — the offered load times the probability your frame survives.”
“Maximizing: take derivative, set to zero. Optimal G = 0.5 (each frame-time has on average half a new frame). Maximum S = 0.5 × e^{-1} = 1/(2e) ≈ 0.184 = 18.4%.”
Ask: “Would you build a highway where only 1 in 5 cars arrives?” This motivates the entire rest of the lecture.

The key takeaway is not the equation — it is the vulnerable window. Every subsequent protocol we study shrinks this window by adding shared state.

Slotted ALOHA (Roberts, 1975)

ALOHA wastes 80% of capacity. The vulnerable window is \(2T\). How would you shrink it?

15 seconds — think about it.

Roberts’ idea: Synchronize all transmitters to a common slot boundary.

Pure ALOHA:     vulnerable window = 2T
                ├──────── 2T ────────┤

Slotted ALOHA:  frames align to slot edges → vulnerable window = T
                ├────── T ──────┤

Frames can only begin at slot edges → no more partial overlaps
Collision = two frames in same slot; otherwise clean
Vulnerable window halves: \(2T \rightarrow T\)

Same formula, smaller window:

\[S = G \cdot e^{-G \cdot (T/T)} = G \cdot e^{-G}, \quad \text{peak at } G=1.0: \textbf{36.8\%}\]

Halving the window doubled throughput. \(G\) didn’t change meaning — only the window did.

Cost: You need a shared clock — one bit of global state (the slot boundary).

The pattern: Halving the vulnerable window doubled throughput. What if you could shrink it further?

Slotted ALOHA’s Problem: Blind Transmission

Kleinrock and Lam (1975) observed two fatal flaws:

Instability under load. As traffic increases, collisions increase → retransmissions increase → more collisions. Positive feedback collapses throughput to zero. The channel has a bistable equilibrium — it randomly flips between working and collapsed states.
Stations ignore available information. In a local network, propagation delay \(\tau\) is tiny (\(\sim 5\) µs). A station can hear other transmissions almost instantly — but ALOHA doesn’t bother to listen.

. . .

The waste: stations transmit into a channel they could have known was busy, if only they had checked.

Slotted ALOHA: "just transmit"

A: [──── TRANSMIT ────]
B:    [──── TRANSMIT ────]  ← collision!
       ↑
       B could have HEARD A
       (τ ≈ 5 µs, T ≈ 1 ms)
       but the protocol doesn't
       ask B to listen first.

       Free information, thrown away.

This slide is the empirical motivation for CSMA. Lam and Kleinrock (1975) proved that ALOHA is inherently unstable — under high load, the system can collapse and stay collapsed. Their solution was dynamic control of transmission probability, which was the first closed-loop MAC.

But the deeper insight came from Kleinrock and Tobagi: in local environments (LANs, building-scale radio), the propagation delay is microseconds while frame transmission takes milliseconds. That means when station A starts transmitting, station B hears it almost immediately — within microseconds. Yet ALOHA protocols never check. They transmit blindly into a channel whose state is freely observable.

This is the intellectual motivation for carrier sensing: stop throwing away free information. If τ << T, then listening before transmitting eliminates most collisions at almost zero cost. The vulnerable window shrinks from T (the whole slot) to τ (the propagation delay) — a 200× improvement on a typical LAN.

“What If You Could Listen First?”

Student exercise: Given what Kleinrock observed — that stations can hear each other almost instantly — redesign the protocol. How does it change?

30 seconds — discuss with your neighbor.

CSMA — Carrier Sense Multiple Access (Kleinrock & Tobagi, 1975):

Carrier sensing = detecting energy on the medium. If your receiver picks up a signal above the noise floor, someone is transmitting — the “carrier” is present. Sense the carrier before you transmit.

Listen before transmitting — if carrier detected (busy), defer
If no carrier (idle), transmit immediately (1-persistent) or with probability \(p\) (\(p\)-persistent)
Vulnerable window shrinks from \(T\) (slot time) to \(\tau\) (propagation delay)
On a LAN: \(\tau \ll T\), so throughput jumps to ~80–90% (vs. 36.8% for Slotted ALOHA)
New shared state: channel busy/idle

. . .

But the assumption can fail:

“If I hear idle, it’s idle everywhere” — wrong.

       ◄──── transmission time T ────►
       ◄τ►  (propagation delay, tiny)

A: idle ├═══════ TRANSMIT ═══════════╡
        ↑ A starts
        │
        ├─τ─┤ A's signal traveling to B
             ↓
B: idle ·····╠═══ TRANSMIT ══════════╡
             ↑ B senses idle (A's signal
               hasn't arrived yet!)
             COLLISION at receiver

  ALOHA vulnerable window:  ◄── 2T ──►
  CSMA  vulnerable window:  ◄τ►
                             ↑ tiny!

On a LAN: \(T \approx 1\) ms, \(\tau \approx 5\) µs → vulnerable window shrinks 200×. Throughput jumps from 36.8% to ~80–90%.

Again, pause and let students propose. Most will say “listen first, then send if quiet.” Push: “What can go wrong? What if two stations both listen, both hear idle, both transmit?” This is the collision vulnerability window for CSMA: it equals the propagation delay between the two farthest stations. If station A transmits and station B is one propagation delay away, B can listen, hear idle (A’s signal hasn’t arrived yet), and transmit — collision.

Kleinrock and Tobagi’s 1975 paper is one of the most cited in networking. The key insight: carrier sensing adds a second dimension to shared state (busy/idle observation), cutting collisions dramatically but not eliminating them. The residual vulnerability window equals propagation delay — the time it takes for A’s signal to reach B.

CSMA/CD: Collision Detection on the Wire (Metcalfe & Boggs, 1976)

On a wire, signals combine at every point. You can hear the medium while transmitting.

Student question: If you could detect a collision during transmission, what would you do differently than CSMA?

Metcalfe’s answer:

Sense channel — if idle, start transmitting
Monitor the wire while sending — compare what you sent vs. what you hear back
Mismatch? → collision detected → abort immediately
Send 48-bit jam signal so all parties detect the collision
Binary exponential backoff: after \(m\)-th collision, pick \(K \in \{0,\ldots,2^m-1\}\), wait \(K \times 512\) bit-times
Retry from step 1

Time ──────────────────────────────►

A: [sense idle][TRANSMIT───✕ abort!][jam][ backoff ]
                           ↑ collision detected
                           │ in ~51.2 µs

B: [sense idle][TRANSMIT───✕ abort!][jam][ backoff ]

   Both detect within 51.2 µs (round-trip
   for max-length Ethernet segment).
   Only a few bytes wasted.

Key numbers: At 10 Mbps, 512 bit-times = 51.2 µs. Collision detected and resolved in microseconds. Utilization: ~95%+.

Why it works: On a wire, all signals combine at every point — a transmitter can hear other signals while transmitting. The wire is a shared bus where interference is directly observable.

But Wireless Cannot Detect Collisions

CSMA/CD works beautifully on a wire. Can we do the same on radio?

10 seconds — what’s different about wireless that might break this?

The root cause is transmission power, which is dictated by the medium:

Guided medium (wire):

Signal confined to conductor → minimal loss
TX power: ~1 mW (0 dBm) suffices
Other stations’ signals on the bus: comparable power
→ Can observe combined signal while transmitting
→ Collision detection works

TX signal:     ████ (0 dBm)
Other signal:  ███  (-3 dBm)
Ratio: ~2×  → easily distinguishable

Unguided medium (radio):

Signal radiates in all directions → \(1/d^2\) to \(1/d^4\) loss
TX power: ~100 mW (20 dBm) needed to reach 50m
Own signal at antenna: -10 dBm
Incoming signal from 50m away: -60 dBm
→ Own signal is 100,000× stronger (50 dB gap)
→ Receiver saturated — deaf while transmitting

TX signal:     ████████████████ (-10 dBm)
Other signal:  ▪                (-60 dBm)
Ratio: 100,000× → invisible

The chain: unguided medium → high TX power needed → own signal drowns incoming → cannot sense while transmitting → collision detection impossible → must avoid, not detect.

This is the critical fork in the road. Ethernet’s wire is a shared bus where all signals combine at every point — a transmitter can easily compare what it is sending against what it hears on the wire. If they differ, collision. Abort in microseconds, waste only a few bytes.

But on a radio, the transmitter’s own signal at its antenna is 100,000 times stronger than any incoming signal. It is like trying to hear someone whisper while you are shouting through a megaphone. No practical radio design (until very recently, with full-duplex research) can cancel the self-interference well enough to detect collisions in real time.

This physics constraint — not a design choice, not a cost tradeoff, but fundamental physics — forced 802.11 to adopt collision avoidance instead of detection. Every subsequent WiFi protocol design decision follows from this anchor.

Concrete CSMA/CD algorithm for reference (Ethernet): (1) NIC creates frame, (2) sense channel — if idle, transmit; if busy, wait. (3) If entire frame sent without detecting collision, done. (4) If collision detected mid-transmission, abort immediately and send a 48-bit jam signal to ensure all parties detect the collision. (5) Binary exponential backoff: after the m-th collision, choose K randomly from {0, 1, …, 2^m − 1}, wait K × 512 bit-times (at 10 Mbps, 512 bit-times = 51.2 µs). At 10 Mbps, collision detection takes ~51.2 µs (the round-trip time for a maximum-length Ethernet segment). This is 1000x faster than CSMA/CA’s 50 ms ACK timeout — the speed difference that explains the throughput gap (~95% vs ~30%).

Framework Checkpoint: The Assumption Cascade

Each generation broke the previous generation’s binding assumption:

Generation	Shared State Added	Throughput	Key Assumption Broken
Pure ALOHA	Nothing	18.4%	“Transmit whenever”
Slotted ALOHA	Slot boundary	36.8%	“Align to slots”
CSMA	Channel busy/idle	~80-90%	“If idle here, idle everywhere”
CSMA/CD (wired)	Collision signal	~95%+	“Can hear own collision”

Wireless can’t do CSMA/CD. The assumption cascade hits a physics wall.

How does wireless close the feedback loop? That’s where we pick up next lecture.

Generative Exercise

What Happens When Physics Changes?

Exercise: When Carrier Sensing Breaks

CSMA works because \(\tau \ll T\) — you hear others almost instantly.

What happens when that breaks?

Step 1: Compute the Ratio

Geostationary satellite: \(\tau = 125\) ms, \(T \approx 5\) ms.

What is \(\tau / T\)? Is carrier sensing useful?

\(\tau / T = 25\). You sense signals that are 25 frame-times old. By the time you hear “busy,” 25 frames could have started and finished. Carrier sensing is useless — you’re back to ALOHA.

Step 2: Design a Protocol

You’re the satellite network engineer. Ground stations can’t hear each other. CSMA is useless.

2 minutes — design a MAC. Discuss with your neighbor.

Guiding questions (reveal one at a time):

Can any distributed protocol beat ALOHA here? Why or why not?

What if the satellite could talk back — what role could it play?

What does this remind you of from cellular networks?

Step 3: The Answer and the Deeper Insight

The natural solution: centralized scheduling. The satellite allocates time slots (TDMA). This is exactly what real satellite systems do (DVB-RCS).

Now the hard question: What if there’s no central controller?

Underwater acoustic network: sound at 1,500 m/s, sensors 1 km apart → \(\tau = 670\) ms, \(T = 100\) ms. \(\tau / T = 6.7\). No infrastructure underwater.

You’re stuck near ALOHA. This is an open research problem.

The insight: The choice between contention and scheduling is not ideological — it is driven by \(\tau / T\). When \(\tau \ll T\), sensing works. When \(\tau \gg T\), centralization becomes necessary. Same physics, same framework.

This is scaffolded as three steps following Winston’s principles:

Step 1: Students compute the ratio themselves. They should viscerally feel that τ/T = 25 makes sensing absurd — you’re looking at ancient history.

Step 2: Students design a protocol. The guiding questions are revealed one at a time to scaffold without giving away the answer. Most groups will arrive at “someone needs to coordinate” → centralized scheduling → TDMA. Push: “Who is the coordinator? The satellite itself? A ground master station?”

Step 3: Reveal the real answer (TDMA/DVB-RCS), then flip the constraint — no controller (underwater). Students should feel the tension: when τ >> T AND there’s no infrastructure, there’s no good answer. ALOHA-like performance is a fundamental limit. This is an active research area.

The τ/T ratio is the key takeaway. It connects backward (CSMA works on LANs because τ/T ≈ 0.005) and forward (L6: WiFi density eventually breaks contention even when τ/T is small, forcing centralization for a different reason — collision probability, not sensing staleness).

Preview: Lecture 6

Wireless can’t detect collisions. So what does it do instead?

CSMA/CA: sense + randomize + ACK — how 802.11 closes the loop
802.11 DCF: the protocol that ran WiFi for 20 years
Hidden terminals: when carrier sensing fails entirely
WiFi evolution: why CSMA/CA eventually broke under density
Cellular: the path that was centralized from day one

Reading: Book Ch 3, §Act 2 (CSMA) through §Act 5 (Cellular Alternative)

Framework Summary: ALOHA to CSMA/CD

Anchor: Shared-medium physics — broadcast, location-dependent sensing

The assumption cascade: each generation adds shared state, shrinks the vulnerable window

Invariant	ALOHA	CSMA	CSMA/CD
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

The wireless fork: unguided medium → high TX power → can’t sense while transmitting → CSMA/CD impossible.

Next: How does wireless close the loop without collision detection?