From exclusive ownership, the cellular invariant answers cascade:

• Coordination → centralized from day one. One entity owns the channel.
• State → can be global. BS knows every device’s location, channel quality, buffer.
• Time → can be precise. BS synchronizes all transmissions to a global clock.
• Interface → can carry rich scheduling information.

Question: The BS knows everything and schedules everything. At what cost?

• CQI (Channel Quality Indicator): SNR summary, every 1–2 ms — cheap, coarse
• CSI (Channel State Information): full channel matrix for beamforming — rich, expensive
• SRS (Sounding Reference Signals): UE sends a predetermined test pattern; BS compares received vs. expected to measure uplink channel distortion

The measurement tax: In massive MIMO (320 MHz, 16 antennas), CSI feedback consumes ~75% of airtime just for measurement overhead.

The scheduler’s global view is purchased at the cost of the very resource it is trying to schedule.

The paradox:

1. Measure frequently → accurate state, but airtime consumed by measurement
2. Measure rarely → save airtime, but state goes stale → wrong scheduling decisions

Neither architecture escapes the cost of coordination. They pay in different currencies.

Question: If a speaker is silent 65% of the time but owns the channel 100% of the time, what is the structural waste?

Hard allocation of a shared resource to bursty demand. Simplicity was the binding constraint in the 1980s.

8 users share one 200 kHz carrier. Each gets a 577 µs slot in a 4.615 ms frame.

• More calls per MHz of spectrum (8 users on 200 kHz vs. 1 user on 30 kHz)
• Hardware: one radio shared across 8 slots (vs. one radio per channel in FDMA)
• Encryption now possible (impossible with analog)

But still hard capacity. All 8 slots full → next caller blocked.

Three CS concepts you already know:

1. XOR encoding — each user’s bits multiplied by a unique spreading code (like XOR with a key)
2. Hash independence — codes are orthogonal (dot product = 0); other users look like noise
3. Redundancy — each bit spread across 128 “chips” (shorter pulses, using 128× more bandwidth than strictly needed). Why pay that cost?

Question: If the codes are orthogonal — perfectly separable — why does adding a user matter at all?

Cocktail party, extended:

• 3 people, 3 languages → murmur is faint, you follow easily
• 30 people → room is noticeably louder
• 100 people → straining to hear your partner

Each speaker adds acoustic energy even if unintelligible. The 128 chips? That’s the receiver’s noise suppression — it averages across 128 samples to extract the signal. Suppresses interference by up to 128×. But with enough speakers, the murmur exceeds even that limit.

Soft capacity: no hard wall. User #61 isn’t blocked — they raise the noise floor for everyone. Degradation is graceful, not binary.

A phone 100m away arrives 10,000× stronger (40 dB) than one 1 km away. Drowns everyone — regardless of codes.

The fix: BS commands each mobile to adjust power 800 times per second, keeping all signals arriving at equal strength.

• Closed-loop power control IS the MAC in CDMA
• Not backoff. Not scheduling. Continuous real-time power management.
• If one user’s loop fails → that user drowns everyone

Today’s question: What is the cost of NOT disaggregating?

The cellular RAN from 1991 to 2009 is a case study in paying that cost — being forced, generation by generation, to separate things that should never have been joined.

No single NodeB can answer this — it only sees its own coverage area. Handover requires a multi-cell view. So 3G placed a controller — the RNC (Radio Network Controller) — above dozens of NodeBs.

The RNC handled: handover coordination, paging (finding which cell a user is in), power control across cells, radio bearer setup.

Question: The RNC sits physically remote (10–20 ms round-trip). For handover — does that matter?

No. Handover decisions tolerate 100 ms. A user crossing cells at 60 km/h gives you seconds of warning. The RNC’s latency was fine for every function it was designed for.

Question: You want to schedule dynamically — give the channel to whoever needs it NOW. The channel’s quality changes every 1–2 ms. The RNC is 10–20 ms away. What breaks?

By the time the RNC’s scheduling decision reaches the NodeB, the channel has faded and the decision is stale.

The RNC didn’t distinguish which decisions were time-critical. Handover (100 ms tolerance) and scheduling (1 ms tolerance) were fused in one remote controller. The cost of not disaggregating: the fastest function is throttled to the speed of the slowest interface.

Now the scheduler sits next to the radio. It re-evaluates every 2 ms — one scheduling cycle. (Cellular calls this cycle the Transmission Time Interval, or TTI.)

• Fast function (who transmits this cycle?) → at the NodeB, where channel info lives
• Slow functions (which BS serves this user?) → stays at RNC
• Decision placement: put decisions where the information lives

The payoff — opportunistic scheduling:

• Give shared resource to whoever has best channel conditions right now
• Ride the peaks of each user’s channel → aggregate throughput increases
• This was impossible when scheduling lived at the RNC

Third disaggregation — the core splits:

• MME (control plane): authentication, session setup, handover
• S-GW + P-GW (user plane): data forwarding
• Scale independently. Voice → just another IP flow (VoLTE).

The rule: a function at timescale X cannot live behind an interface imposing latency > X.

The cost of NOT splitting: fastest function throttled to speed of slowest interface.

The universal architecture:

Schedule everything you can. Contend only for discovery.

• Cellular: RACH (slotted ALOHA) for device discovery → scheduling for everything else
• WiFi: CSMA/CA for initial association → OFDMA for data

Identical constraints → identical invariant answers, regardless of starting point.

Faster loop → tighter coordination → higher utilization. But faster loop → heavier measurement tax.

Convergence is not “centralization wins.” It is: centralization trades collision overhead for measurement overhead. At scale, that tradeoff is better.

\(P(\text{success}) = 26 \times 0.125 \times (0.875)^{25} \approx 0.106\) — 10% of slots succeed. Density is NOT the problem.

But the robot’s haptic packet enters the same FIFO queue as heartbeat readings and video chunks. Worst case: tablet TXOP (5 ms) + collisions + backoff → 5–15 ms delay. The 10 ms deadline breaks — not from saturation, but because the MAC has no concept of urgency.

The binding constraint shifted: density → service diversity. 500 ms vs. 50 ms vs. 10 ms on one queue. This is exactly the problem 5G faces at network scale — and why we need network slicing (next lecture).

But centralization creates a new monolith. The scheduler is locked in proprietary hardware. One vendor. One algorithm. One box.

L9: How do you crack that monolith open?

• CU / DU / RU — disaggregate the base station along timescale boundaries
• O-RAN — open the interfaces (the cellular SDN moment)
• Network Slicing — serve contradictory requirements on shared infrastructure