Breaking the Contention Ceiling

CS176C — Advanced Topics in Internet Computing

Arpit Gupta

2026-04-23

Where We Left Off

Last lecture: we proved CSMA/CA hits a structural throughput ceiling under density.

Bianchi (2000): ~30% utilization at \(n \approx 50\) — no BEB tuning fixes it
State grew at every step (ALOHA → CSMA → CSMA/CA → DCF)
Coordination stayed distributed the entire time

The ceiling is not a state limit. It is a coordination limit.

Today’s question: WiFi shipped this protocol in 1997 and survived 20 years. What did it actually do about the ceiling — and why did it take until 2021 to break it?

The binding constraint: WiFi operates in unlicensed spectrum. No entity owns the channel. From that single fact, the invariant answers cascade: Coordination → distributed, State → local, Time → coarse, Interface → minimal. CSMA/CA is not a design choice — it is the only viable MAC under these constraints.

Part 1: WiFi’s Co-Evolution

Making the Pipe Wider — and Discovering It Wasn’t Enough

The Pressure Chain

WiFi’s first instinct: if CSMA/CA wastes most of the channel, make the channel so large that even 30% is enough.

Three simultaneous advances, each enabling the next:

1. Modulation — more bits per symbol

802.11b (1999): 11 Mbps via Complementary Code Keying (CCK)
802.11a (1999): 54 Mbps via Orthogonal Frequency Division Multiplexing (OFDM) — splits the channel into many narrow subcarriers, each carrying data; robust against indoor multipath
Modulation up to 64-Quadrature Amplitude Modulation (QAM): 6 bits per symbol
802.11ac (2013): 256-QAM → 8 bits/symbol; 802.11ax (2021): 1024-QAM → 10 bits/symbol

2. Antennas — Multiple-Input Multiple-Output (MIMO)

Spatial diversity: combine signal copies from different paths → better Signal-to-Noise Ratio (SNR) without more power
Spatial multiplexing: 4 antennas → up to 4 independent data streams simultaneously

3. Channel width — 20 MHz → 40 → 80 → 160 MHz

Multiply together: 8 streams × 256-QAM × 160 MHz = 6.9 Gbps peak PHY rate (802.11ac)

This slide compresses the WiFi PHY evolution into one logical chain. Don’t dwell on any single generation — the point is the pattern: all three advances work in concert, and each alone is insufficient. Higher modulation without MIMO can’t meet the SNR bar. MIMO without higher modulation wastes spatial capacity. Wider channels without MIMO create more noise to overcome.

Key numbers to say aloud: 64-QAM needs ~26 dB SNR. Binary Phase Shift Keying (BPSK) needs only ~7 dB. As you walk away from the AP, SNR drops, and the system drops to simpler modulation — this is rate adaptation. The “speed” on your WiFi status bar is the output of a per-packet feedback loop varying moment to moment.

The 6.9 Gbps number is the dramatic payoff. But preview the paradox: “6.9 Gbps is the engineering achievement. What happens when CSMA/CA tries to use it?”

MU-MIMO: The Promise

Downlink Multi-User MIMO (MU-MIMO) in 802.11ac: the AP uses beamforming — shaping the antenna pattern to focus energy toward specific clients — to serve up to 4 users simultaneously on different spatial beams.

The question: if the AP can serve 4 users in parallel, does throughput quadruple?

MU-MIMO: The Half-Victory

Partly. 802.11ac cracked the downlink: the AP wins a CSMA/CA contention round, then beamforms to 4 users in parallel during that single Transmit Opportunity.

Downlink delivery is genuinely parallel. 802.11ac DID break the one-user-at-a-time paradigm — for the downlink.

But: the uplink is untouched — every client still contends one at a time. And the AP itself must fight those same clients just to get a turn.

MU-MIMO: What Remains Broken

Downlink parallelism exists, but it’s gated by two bottlenecks:

1. Serial access. The AP must win CSMA/CA contention against every client before starting its MU-MIMO burst. At n=250, winning is the problem.

2. One direction only. Uplink is still one client at a time. The 6.9 Gbps peak assumes the AP has the channel — getting the channel is still a contention problem.

Even the downlink win is taxed: MU-MIMO requires a channel-sounding handshake before every burst — overhead that can negate the parallelism gain for small packets.

The Overhead Tax

Your pre-lecture exercise used 100 µs as protocol overhead. The real number — including average backoff — is closer to 160 µs. That makes the picture worse.

Component	Value	Why it’s fixed
DIFS (DCF Interframe Space)	34 µs	Radio turnaround physics
Average backoff	67.5 µs	\(CW_{\min}=15\), avg = 7.5 slots × 9 µs/slot
PHY preamble	20 µs	OFDM training symbols
SIFS (Short Interframe Space)	16 µs	Radio RX-to-TX switch
ACK at 6 Mbps basic rate	~24 µs	14 bytes + preamble at lowest mandatory rate
Total	~160 µs	None scale with PHY rate improvements

What happens as the PHY gets faster?

PHY rate	Frame TX (1500 B)	Overhead	Data fraction
54 Mbps	222 µs	160 µs	58%
600 Mbps	20 µs	160 µs	11%
6.9 Gbps	1.74 µs	160 µs	1.1%

The PHY engineers gave us 6.9 Gbps. The MAC wastes 99% of it on protocol gaps. Making the pipe wider made the overhead ratio catastrophically worse.

Walk through the corrected-overhead table carefully. The 100 µs from the exercise was a deliberate simplification — it omits average backoff (67.5 µs at CW_min=15 with 9 µs slots). The precise breakdown totals ~160 µs (or ~162 µs if you’re being exact).

Key distinction to draw explicitly: the overhead tax hits even a SINGLE station on an EMPTY channel. It is distinct from the contention ceiling (which comes from collision probability growing with n). Both limit throughput. Both stem from the same root: CSMA/CA’s per-frame protocol structure was designed when frames took milliseconds, not microseconds.

The data-fraction table is the emotional climax. At 54 Mbps, 58% is “not great but workable.” At 600 Mbps, only 11% — five-sixths of airtime is protocol overhead. At 6.9 Gbps, 1.1% — 99% waste. The data shrunk to a sliver inside a vast protocol envelope.

Verbal fence: “overhead tax ≠ contention ceiling. The overhead tax is a per-frame structural cost. The contention ceiling is a per-station scaling cost. Both are real, both are bad, and they compound.”

Aggregation: The MAC’s First Response

The overhead tax is 160 µs per frame. What if you could pay it once for many frames?

802.11n introduced Aggregated MAC Protocol Data Unit (A-MPDU): bundle up to 64 frames under a single PHY preamble.

Without aggregation: 1 preamble, 1 DIFS, 1 backoff, 1 ACK → per frame
With aggregation: 1 preamble, 1 DIFS, 1 backoff → 64 frames → 1 Block ACK

Block ACK: a bitmap where each bit marks one frame as received or lost — selective retransmission of only failed frames.

Transmit Opportunity (TXOP) (from earlier 802.11e amendment): the winner holds the channel for a burst duration instead of re-contending after every frame.

Result: per-frame overhead drops from 160 µs to ~2.5 µs. Data fraction recovers.

In the framework: aggregation is a Time invariant fix — it changes the effective timescale of the protocol. But Coordination remains distributed. State remains local. The binding constraint hasn’t changed. The ceiling is still there; aggregation raised the floor beneath it.

Walk through the mechanism: A-MPDU takes up to 64 individual frames, each with its own MAC header, and wraps them under a single PHY preamble (the 20 µs training sequence). One contention event, then a burst of data. The receiver confirms with a Block ACK — a compact bitmap rather than 64 individual acknowledgments. Together with TXOP (which lets the winner hold the channel for a burst duration), aggregation amortizes the 160 µs overhead across many payloads.

This was enough for 802.11n/ac era — homes, small offices, coffee shops with 5-10 devices. Aggregation gave MIMO a way to actually deliver its PHY gains.

The framework reading is important: aggregation changes WHEN you pay the overhead (once per burst vs. once per frame), not WHO decides. It’s a time-domain fix, not a coordination fix. The ceiling persists.

Transition: “Aggregation fixed the per-frame overhead. It did not fix contention. Let’s quantify why.”

Density Breaks the Fix

Aggregation fixed per-frame overhead. It did not fix contention. Let’s quantify why.

Define \(\tau\) = probability a station transmits in any given slot. Not “per attempt” — per slot. If we freeze time at a random slot boundary, \(\tau\) is the chance station \(i\)’s backoff counter is at zero.

If every station is at \(CW_{\min}\) (no collisions yet): \(\tau \approx \frac{2}{CW_{\min} + 1}\). For \(CW_{\min} = 15\): \(\tau = 0.125\).

In this one slot, what’s the probability exactly one station transmits?

\[P(\text{success}) = n \cdot \tau \cdot (1 - \tau)^{n-1}\]

Pick one transmitter (\(\tau\)), require all other \(n-1\) to stay silent (\((1-\tau)^{n-1}\)), any of \(n\) could be the one.

\(n = 10\): \(P \approx\) 0.40 — 40% of slots succeed. Workable.

\(n = 50\): \(P \approx\) 0.009 — less than 1%.

\(n = 250\): \(P \approx\) 0. Virtually no successful slot.

Caveat: we assumed \(\tau = 0.125\) — all stations at \(CW_{\min}\), no collisions yet. At high \(n\), collisions are constant, BEB grows CW, and the real stationary \(\tau\) is much smaller. Our numbers are pessimistic. But Bianchi’s full model (which solves for the true equilibrium \(\tau\)) still shows throughput dropping to 40-55% at \(n=50\) and falling further at higher density. The collapse is real — our approximation just makes it look worse than steady state.

Walk through the combinatorics carefully:

tau = per-slot transmission probability (stationary, NOT “per attempt”). Think of it as: at any random slot boundary, what fraction of stations happen to have their backoff counter at zero?

tau = 2/(CW_min+1) is the LOW-LOAD approximation — valid when collisions are rare and all stations stay near CW_min. At high n, BEB pushes the average CW far above CW_min, making the true stationary tau much smaller.

P(success) = n * tau * (1-tau)^(n-1) is a snapshot of ONE slot: exactly one of n stations transmits, all others stay silent. It’s not about multiple attempts — it’s the probability that this specific slot produces a successful transmission.

The numbers at n=50 and n=250 using tau=0.125 are deliberately pessimistic. They show what happens if everyone were still at CW_min. In reality, BEB grows the average window, so the real tau is lower and collisions are somewhat less catastrophic. But Bianchi’s full model (which solves coupled equations for the equilibrium tau and collision probability p) still shows sharp throughput degradation above ~30 stations. The curve’s shape is structural.

AP asymmetry: the AP carries 10-100x more data but contends on equal terms. One station, one slot, same probability. The AP holding 95% of data gets 1/(n+1) of access. This is the setup for WiFox and the coordination shift.

WiFox: A Step Toward Centralization

This wasn’t just theory — it was observable in the wild.

In 2012, studying WiFi in packed lecture halls and conferences: pages wouldn’t load, video buffered endlessly — even at hundreds of Mbps PHY rate.

Root cause: the AP was starving for channel access. Massive downlink queue, but it could only transmit when it won contention against hundreds of clients — each with almost no data but equal contention rights.

The protocol’s fairness, designed for 5 devices, was the source of failure at 500.

WiFox (ACM CoNEXT 2012): software-only fix. Dynamically shorten the AP’s backoff window when its queue depth grows — giving the AP priority over clients.

400-700% improvement in downlink throughput
30-40% reduction in response time
Deployable as firmware update — no hardware or client changes

In the framework: WiFox was a partial Coordination shift. We gave the AP privileged access — it no longer competed on equal terms. We broke CSMA/CA’s per-station fairness because that fairness was the source of failure.

WiFox was a step toward centralization. The full step came with 802.11ax.

This is a personal research story — Arpit was a grad student at NC State when WiFox was developed and published at ACM CoNEXT 2012. The work traced the root cause of WiFi failure in dense environments to exactly the AP downlink starvation described on the previous slide.

WiFox monitored the AP’s queue depth. When the buffer was growing — a signal that downlink was starving — it shortened the AP’s backoff window, effectively giving it priority. Simple, deployable, no hardware changes.

In the framework: WiFox broke CSMA/CA’s per-station fairness. It gave the AP a shorter contention window, meaning it was no longer “one station among many.” This is a partial coordination shift — the AP is now privileged, but it still contends. The full centralization came with 802.11ax, where the AP becomes a scheduler and contention is eliminated entirely on the data path.

The key observation for students: when the binding constraint shifted from “no authority” to “coordination cost exceeds capacity,” fairness itself became the enemy. WiFox recognized this before the standards did.

802.11ax: WiFi Centralizes

802.11ax (WiFi 6, 2021) changed the MAC architecture fundamentally.

Instead of stations contending for the whole channel one at a time, the AP divides it into Resource Units (RUs) — time-frequency slots assigned to individual clients.

How many RUs? Each RU = 26 Orthogonal Frequency Division Multiple Access (OFDMA) subcarriers ≈ 2 MHz. In a 160 MHz channel: \(160/2 = 80\) naively, but 74 RUs in practice. The missing 6 RUs’ worth of subcarriers go to guard bands, DC null subcarriers, and inter-RU gaps.

The mechanism: Trigger Frame.

A specialized AP broadcast carrying per-client metadata — which RU, what modulation, how much power, how long. Clients cannot transmit on OFDMA RUs without a Trigger Frame. All assigned clients begin transmitting exactly one SIFS (16 µs) later.

Multiple clients transmit simultaneously on non-overlapping frequency slices. Zero contention on the data path.

Follow the logic: Coordination must be centralized → someone needs a global view → that’s the AP → the AP must partition resources deterministically → that structure is OFDMA. WiFi didn’t borrow OFDMA from cellular Long-Term Evolution (LTE) out of convenience — the constraint shift forced the same invariant answers.

This is the climax of the WiFi arc. Walk through the Trigger Frame mechanism carefully:

The AP broadcasts a Trigger Frame — a downlink control frame containing per-client scheduling information: which client gets which RU, what modulation and coding scheme, how much transmit power, how long the window lasts.
Clients CANNOT transmit on OFDMA RUs without receiving a Trigger Frame first — that’s the centralization made concrete.
Upon receiving it, all assigned clients begin transmitting exactly one SIFS (16 µs) later, synchronized in time and frequency across their respective RUs.
Multiple clients transmit simultaneously on non-overlapping frequency slices. Zero contention. Zero collision on the data path.

The 74 RUs number: 160 MHz / 2 MHz per RU = 80 naively, but guard bands at channel edges (prevent spectral leakage), DC null subcarriers at center (protect against hardware impairments), and inter-RU gaps (prevent adjacent-RU interference) consume 6 RUs’ worth. So 74 parallel users per Trigger Frame.

CSMA/CA is retained for exactly two purposes: (1) serving legacy clients that don’t speak 802.11ax, and (2) initial association — when a new device joins and the AP doesn’t yet know it exists.

What WiFi Lost — and Why 20 Years

What WiFi lost: any station can transmit whenever it wants, no permission needed, no infrastructure required. Under 802.11ax, the high-throughput path requires an AP scheduler. No AP, no OFDMA. Ad-hoc and mesh networks fall back to CSMA/CA.

What WiFi gained: channel utilization above 70%, multi-user parallel transmission, deterministic latency, elimination of hidden terminals on the OFDMA path.

Why did centralization take twenty years? Three forces:

1. No operator. Cellular has a carrier that owns spectrum and runs the scheduler. WiFi has no single entity — the AP-as-scheduler only became obvious when APs became ubiquitous infrastructure.

2. Backward compatibility. An 802.11ax AP must still serve an 802.11b client from 1999. You cannot remove CSMA/CA — only layer scheduling on top.

3. Deployment context. Before ~2012, a home AP served 3-5 devices. The ceiling was theoretical, not lived. It took lecture halls, airports, and stadiums to turn the math into a visible problem. Technology adoption follows deployment pressure, not theoretical possibility.

Two parts to this slide. First: what the invariant shift cost. WiFi’s defining character was permissionless access — any device, any time, no infrastructure needed. 802.11ax broke that for the high-throughput path. The framework reading: State moved from local-only to AP-aggregated. Coordination moved from fully distributed to centralized. The AP became a base station.

Second: why 20 years. The three forces are non-technical. Engineers could have centralized in 2005. The market didn’t need it until 2015. This is important for students to understand — system evolution is driven by deployment pressure, not by theoretical possibility. The contention ceiling was known since Bianchi (2000), but the environments that triggered it (hundreds of devices per AP) didn’t exist until smartphones proliferated.

The backward compatibility force is particularly constraining: every new WiFi generation must interoperate with every previous one. You can’t remove CSMA/CA; you can only add scheduling on top. This adds complexity and limits the scheduler’s freedom.

Part 2: The Other Path

Cellular — Licensed Spectrum, Centralized from Day One

Licensed Spectrum: The Mirror Constraint

WiFi’s binding constraint: unlicensed spectrum → distributed coordination forced.

Cellular’s binding constraint: licensed spectrum — exclusive operator ownership, bought at auction for billions → centralized coordination from day one.

The base station is the sole authority over every transmission. There is no contention for data, ever.

State can be global (the BS knows every registered device)
Time can be precise (the BS synchronizes all transmissions)
Interface can carry rich scheduling information

Cellular’s evolution question was never “when to centralize” — it was “how to schedule efficiently as traffic changed from voice to data.”

Voice → Data: the scheduling challenge

Frequency Division Multiple Access (FDMA), 1G (1980s): 30 kHz channels, one per call, ~60 per cell. Hard capacity — call 61 is blocked. Average voice activity ~35%, so two-thirds of spectrum carries silence.

Time Division Multiple Access (TDMA) / GSM, 2G: 8 time slots per 200 kHz carrier, 576.9 µs each in a 4.615 ms frame. More users, better codecs. Still hard capacity.

Code Division Multiple Access (CDMA), 3G: all users share the same frequency, separated by orthogonal spreading codes. Adding a user raises the noise floor — soft capacity instead of hard. But power control becomes the MAC: the base station adjusts each mobile’s transmit power 800 times/second to prevent the near-far problem (a phone 100 m away arrives 10,000× stronger than one 1 km away).

This slide compresses the cellular arc into one view. The key pedagogical move: cellular started where WiFi ended up (centralized), so the evolution question is different. WiFi asked “when to centralize?” Cellular asked “how to schedule better?”

FDMA: Think of it as a hotel with fixed rooms — one per guest, no sharing. If the hotel is full, the next guest is turned away. And guests sleep 65% of the time (voice activity factor ~35%), so most rooms are wasted.

TDMA/GSM: same hotel, but now each room has 8 time-share beds. More guests fit, but still hard capacity.

CDMA: all guests share one big room, separated by noise-canceling headphones (spreading codes). “Orthogonal” means when you multiply the combined signal by a specific user’s code, all other users average to near-zero. Adding a guest raises background noise for everyone — soft degradation rather than hard blocking. But power control is critical: without it, a nearby phone drowns out a distant one entirely (near-far problem). The BS adjusts each phone’s transmit power 800 times/second.

The soft-capacity model was the conceptual precursor to data scheduling: if resources can be shared dynamically rather than allocated in fixed blocks, you can match allocation to demand on short timescales.

All three systems — FDMA, TDMA, CDMA — were designed for voice: steady, symmetric traffic. What happens when traffic becomes bursty? That’s the next slide.

From Voice to Data: The Scheduling Bridge

Voice: steady, symmetric — well-served by dedicated circuits.

Data: loading a webpage involves a few hundred milliseconds of transfer, then seconds of idle reading. Under the circuit model, you hold a dedicated channel the entire time — carrying data for a fraction of a second, then silence for ten seconds.

What scheduling concept bridges circuit-switched voice to packet-switched data?

High Speed Downlink Packet Access (HSDPA), deployed ~2005: replaced dedicated per-user channels with a shared pool scheduled every 2 milliseconds — one Transmission Time Interval (TTI).

Each TTI, the scheduler picks users with the best channel conditions
Channel conditions reported via Channel Quality Indicator (CQI) feedback — the cellular answer to the State invariant
TTI = the cellular answer to the Time invariant: how often does the scheduler re-evaluate?

HSDPA proved the concept: per-TTI packet scheduling works at millisecond timescales. LTE built its entire architecture around OFDMA — the same OFDMA that WiFi borrowed in 802.11ax, twelve years later.

HSDPA is the bridge between 3G circuit-switched voice and 4G packet-switched data. It kept CDMA’s air interface but replaced dedicated per-user channels with a shared downlink pool that the base station schedules every 2 ms.

The key insight: instead of assigning a fixed channel to each user (who mostly wastes it), give the shared pool to whoever can use it best RIGHT NOW. CQI feedback tells the BS which users have the best channel conditions at any given moment. The scheduler exploits this — giving resources to users with good channels (opportunistic scheduling) increases total throughput.

TTI is a Time invariant example: it’s the clock period of the scheduling loop. Every 2 ms, the BS observes (CQI), decides (scheduler), acts (allocates resources). Faster TTI → tighter feedback → better utilization. LTE kept the OFDMA structure and pushed TTI down to 1 ms. 5G NR pushed it further.

Connect to WiFi: 802.11ax’s Trigger Frame is conceptually the same as LTE’s downlink scheduling grant — a centralized control message that tells each client exactly when and where to transmit. WiFi and cellular converged on the same mechanism because the same constraints forced the same answers.

The Universal Bootstrap

GSM is fully centralized — zero contention for data. But there’s one moment when centralization fails.

When? Think about it. When can’t the base station schedule a device?

When a brand-new device powers on. The BS doesn’t know it exists. It can’t schedule what it hasn’t discovered.

Solution: Random Access Channel (RACH) — the device announces itself using slotted ALOHA. The 1972 protocol, running inside a 1991 TDMA system.

Could you eliminate RACH entirely? Could you build a system with zero contention at every level?

No. Any system with dynamic membership needs at least one contention-based moment: the discovery moment. You can centralize everything after discovery, but discovery itself is irreducibly contention-based.

802.11ax has the same pattern: OFDMA for data, CSMA/CA for initial association. Contention is the tax you pay for letting new participants join without prior arrangement.

This is one of the most satisfying observations in the entire landscape. RACH is slotted ALOHA — the simplest protocol from L5 — running inside the most sophisticated cellular system. It exists because there is no alternative for the bootstrap problem.

Walk students through: a new phone powers on. The BS has no record of it. There’s no scheduling entry. The phone must announce itself somehow. It does so by transmitting in a designated RACH slot, hoping no other phone picked the same slot. If two phones collide, they back off and retry. This is literally slotted ALOHA.

The generalization: ANY system with dynamic membership — WiFi, cellular, Bluetooth, satellite — needs at least one contention-based moment for discovery. You can centralize everything after that, but the bootstrap is irreducibly distributed.

802.11ax parallel: OFDMA for data (centralized), CSMA/CA for association (contention). Same pattern. The universal architecture is: schedule everything you can, contend only for discovery.

Feedback Loop Speed: What Unifies the Landscape

Every wireless MAC is a control loop: measure → allocate → transmit → measure outcome → reallocate.

How fast can that loop run? This determines the architecture’s ceiling.

System	What it measures	Loop period	Coordination
CSMA/CA	Carrier sense (local, binary)	~12 ms	Distributed
GSM TDMA	Slot assignment	~4.6 ms	Centralized
CDMA power control	Received power per user	~1.25 ms	Centralized
LTE OFDMA	CQI per user per RB	~1 ms	Centralized
802.11ax OFDMA	Per-client feedback	~1-5 ms	Centralized

Faster loop → tighter coordination → higher utilization.

CSMA/CA had the slowest loop in the landscape. That’s the structural reason contention broke — and why centralization, with its faster and richer feedback, was the destination for both WiFi and cellular.

This table unifies the entire landscape through one lens: loop speed. Every wireless MAC is fundamentally a feedback control system. The speed of observation-decision-action determines how well the system can track changing conditions and avoid waste.

CSMA/CA’s loop is the slowest because it measures locally (carrier sense at the transmitter) and gets feedback only after the entire frame exchange (ACK or timeout). ~12 ms at legacy rates.

GSM: the BS assigns slots in a 4.615 ms frame structure. Slot assignments change frame-to-frame.

CDMA: power control runs at 800 Hz (1.25 ms). This is fast enough to track vehicular fading — a car moving at highway speed causes signal fluctuations on the order of milliseconds.

LTE: CQI reports let the BS reallocate resource blocks every 1 ms TTI. Each device reports how good its channel is on each frequency slice.

802.11ax: comparable to LTE, with Trigger Frames enabling per-client scheduling at millisecond timescales.

The insight: this is different from the overhead tax (which is about per-frame airtime waste). Loop speed is about how quickly the system adapts to changing conditions. CSMA/CA has both the highest overhead tax AND the slowest loop — a double penalty.

Part 3: Convergence

Two Paths, One Destination

Convergence: Each Side Borrowed the Other’s Technique

After thirty years of divergence, WiFi and cellular arrived at the same architecture.

WiFi started fully distributed → fought through 20 years of PHY improvements and MAC patches → centralized under density pressure → borrowed OFDMA from LTE.

Cellular started centralized from day one → entered unlicensed 5 GHz band through Licensed-Assisted Access (LAA) in 2016 → couldn’t schedule spectrum it didn’t own → adopted Listen Before Talk (LBT) — carrier sensing, the WiFi mechanism.

Each side borrowed the other’s technique for the exact reason the other had adopted it originally.

WiFi borrowed scheduling because density demanded it
Cellular borrowed contention because shared spectrum demanded it

Both retain contention for exactly one purpose: bootstrap.

Cellular: RACH for device discovery
WiFi: CSMA/CA for initial association

The universal architecture: schedule everything you can, contend only for discovery. The destination was determined by physics and density. The path depended on where you started.

This is the intellectual payoff of the entire lecture. WiFi and cellular converged because they faced the same physics. WiFi started distributed and moved toward centralization as density demanded it. Cellular started centralized and moved toward contention when it entered shared spectrum.

LAA (Licensed-Assisted Access, 3GPP Release 13, 2016): cellular operators wanted to use unlicensed 5 GHz spectrum to supplement their licensed bands. But in unlicensed spectrum, they can’t schedule — other devices (WiFi APs, other operators) are transmitting too. So LTE adopted Listen Before Talk — sense the channel before transmitting, back off if busy. This is carrier sensing, the core WiFi mechanism. Cellular had to learn contention because the spectrum constraint demanded it.

The bootstrap observation generalizes: any system with dynamic membership needs contention for discovery. You can centralize everything after that. This is true of WiFi, cellular, Bluetooth, satellite — any shared-medium system.

The framework prediction: identical constraints force identical invariant answers, regardless of starting point. WiFi and cellular arrived at the same State (AP/BS-aggregated), same Time (millisecond scheduling loops), same Coordination (centralized), same Interface (OFDMA resource allocation).

In-Class Exercise: CSMA/CA vs. OFDMA at Scale

Setup: You’re an AP in a stadium section. 250 phones. 160 MHz channel. PHY rate: 600 Mbps.

Part 1 — Using the formulas from earlier in the lecture:

\(\tau = 2/(CW_{\min} + 1) = 0.125\) and \(P(\text{success}) = n \cdot \tau \cdot (1 - \tau)^{n-1}\)

Compute P(success) at \(n = 10\), \(n = 50\), \(n = 250\). (3 minutes with your neighbor.)

\(n = 10\): 0.40 — 40% of slots succeed.

\(n = 50\): 0.009 — less than 1%.

\(n = 250\): ≈ 0 — virtually no successful slot.

Give students 3 minutes to compute. The math is straightforward — the point is the shape of the collapse, not the calculation itself.

Walk through the n=50 computation explicitly: ln(0.875) = -0.1335, so 49 x (-0.1335) = -6.54, e^(-6.54) = 0.00144. P = 6.25 x 0.00144 = 0.009.

For n=250: (0.875)^249 = e^(249 x -0.1335) = e^(-33.3) = 4e-15. Multiplied by 250 x 0.125 = 31.25, still essentially zero.

Caveat for students who ask: tau = 0.125 assumes all stations at CW_min (no prior collisions). At n=250, BEB would have grown the average CW far beyond 15, making the real stationary tau much smaller. These numbers are pessimistic — they show what one slot looks like if everyone were at CW_min simultaneously. Bianchi’s full model solves for the equilibrium tau and still shows throughput degradation at scale, just less catastrophic than this low-load approximation suggests.

Exercise Part 2: OFDMA Parallelism

Same 160 MHz channel. Now with 802.11ax OFDMA.

Each RU = 26 subcarriers ≈ 2 MHz. How many parallel RUs?

\(160 / 2 = 80\) naively → 74 RUs in practice (guard bands, DC null, inter-RU gaps).

Each Trigger Frame serves up to 74 users simultaneously — zero contention, zero collision.

Side by side:

	CSMA/CA (\(n = 250\))	OFDMA
Users served per access	1 (if it succeeds — almost never)	Up to 74 (every time)
Contention	Catastrophic	None
Per-user bandwidth	Full 160 MHz (winner-take-all)	~2 MHz per RU

The tradeoff: under CSMA/CA, the winner gets the full channel. Under OFDMA, each user gets ~2 MHz. At 5 devices, CSMA/CA’s winner-take-all is fine. At 250, 2 MHz of guaranteed collision-free access beats fighting over 160 MHz and getting nothing.

Density is what flips the tradeoff. Which invariant changed? Coordination — distributed → centralized. What forced it? Density shifted the binding constraint from “no authority” to “coordination cost exceeds capacity.”

The side-by-side comparison is the climax of the exercise. Under CSMA/CA at n=250, virtually no slot produces a successful transmission — the system is choked by collisions. Under OFDMA, every single Trigger Frame successfully serves up to 74 users in parallel.

The tradeoff is real: CSMA/CA gives the winner the full 160 MHz. OFDMA with 74 RUs gives each user only ~2 MHz. In sparse scenarios (5 devices on a home AP), CSMA/CA’s “winner takes all” is actually fine, and OFDMA’s scheduling overhead may not be worth it. But at 250 users, the comparison is 2 MHz guaranteed vs. 160 MHz theoretical with zero actual throughput.

The insight students should articulate: the difference isn’t incremental, it’s structural. CSMA/CA at n=250 produces effectively zero throughput because distributed coordination collapses. OFDMA serves everyone because centralized scheduling eliminates contention entirely. Could you have predicted the need for OFDMA from the constraint shift alone? Yes — once coordination must be centralized, the AP needs a global view and must partition resources deterministically. That structure is OFDMA.

The Invariant Shift: 1997 → 2021

Invariant	CSMA/CA + DCF (1997)	802.11ax OFDMA (2021)
State	Local only — each station tracks its own backoff counter	AP-aggregated — per-client channel quality, buffer status, scheduling decisions
Time	Frame-time-bounded feedback (~12 ms)	Scheduling-cycle feedback (~1-5 ms)
Coordination	Fully distributed — each station decides independently	Centralized — AP schedules all transmissions
Interface	802.11 frames (ACK, RTS/CTS, NAV)	802.11 frames + Trigger Frame + OFDMA RU assignments

The pattern:

State, Time, Coordination all shifted. Breaking Coordination is what broke the ceiling.
Interface didn’t change — 802.11ax still speaks 802.11 frames. Backward compatibility locked the interface, and it’s the reason the transition took twenty years, not five.

The lesson is structural: when density shifts the binding constraint, the invariant answers converge — regardless of starting point. WiFi and cellular arrived at the same State, same Time, same Coordination, same Interface. Two different paths, one destination. The framework predicted it.

But centralization creates its own problem: the scheduler is a single point of complexity — and of failure. How do you build, scale, and open up that scheduler? That’s the infrastructure question, and it’s next.

This is the closing framework slide. Walk through each row:

State: from “each station knows only its own backoff counter and NAV” to “the AP holds per-client channel quality reports, buffer status, and makes global scheduling decisions.”

Time: from frame-time-bounded feedback (~12 ms per attempt at legacy rates) to scheduling-cycle feedback (~1-5 ms Trigger Frame intervals). The loop got 5-10x faster.

Coordination: the invariant that stayed fixed through the entire evolution from ALOHA to DCF — from Pure ALOHA (1970) through Slotted ALOHA, CSMA, CSMA/CD, CSMA/CA, to DCF (1997) — FINALLY changed. Fully distributed to centralized. This is the invariant that broke the ceiling. It held for 27 years and five protocol generations.

Interface: didn’t change. 802.11ax still speaks 802.11 frames. This is the backward compatibility constraint — and it’s why the transition took 20 years. You can’t break the interface without breaking every existing device.

The cliffhanger: centralization solves the contention problem but creates a new one. The scheduler is now a single point of complexity. How do you build it? Scale it? Open it up so third parties can innovate on scheduling algorithms? That’s 5G, O-RAN, and network slicing — next lecture.