5  Wireless Infrastructure — RAN Disaggregation and Cellular Architecture

---

5.1 The Anchor: Licensed Spectrum with Exclusive Operator Ownership

Cellular networks sit at the opposite end of the coordination spectrum from WiFi. Where WiFi is born of unlicensed spectrum and distributed contention, cellular is born of licensed spectrum with exclusive operator ownership. A mobile operator buys a slice of spectrum — 700 MHz, 2.1 GHz, 3.5 GHz, 28 GHz — and holds monopoly rights to transmit in that band within a geographic region. This monopoly on the air interface has a profound architectural consequence: a single authority can schedule every transmission. Every station receives an explicit schedule for when to transmit; the base station tells it when to send and on which resource blocks. This exclusive ownership is the binding constraint every cellular architecture inherits.

But exclusive ownership creates a second problem: users move. A subscriber walks, drives, or rides between cells in seconds. Each move requires transferring the session — the radio link, the IP address anchor, the security context, the billing state — from one base station to another without dropping the call. Mobility forces coordination across base stations, coordination that exceeds any single scheduler’s scope.

These two constraints — centralized scheduling per cell + coordination across cells — are inherited and persist across all designs. They create the binding constraint for every cellular infrastructure design: scheduled access with cross-cell session transfer.

The NGMN Alliance, gathering operators in 2015 to define what 5G must become, framed the architectural demand that followed:

“5G systems will need to support an unprecedented scaling in traffic volume, number of connections, and service diversity — from mission-critical machine communications at one extreme to immersive media at the other — on a unified, programmable infrastructure.” — NGMN Alliance, 2015 (NGMN Alliance 2015)

Writing seven years later as O-RAN took shape, Polese et al. formalized what 5G’s architecture evolution had been about:

“Open, programmable, and virtualized networks represent a fundamental departure from the closed, monolithic RAN designs of prior generations — the RAN is no longer a black box but a disaggregated, software-defined platform.” — Polese et al., 2023 (Polese et al. 2023)

That framing was retrospective. The GSM architects of 1991 set out to build a voice service, with programmability far from view. They built a monolith because the service was voice, the hardware was custom silicon, and the only constraint that mattered was delivering a reliable phone call. The disaggregation story unfolded across four decades, one generation at a time.

The infrastructure must continuously answer four decision problems:

1. Where to place scheduling logic — monolithic or distributed? Keep it all in one box next to the antenna, or split it across a centralized pool and a remote radio head?
2. How to hand off a connection across base stations without dropping it? Transfer radio link, session state, and IP anchor in under 150 ms while the user is moving at 120 km/h.
3. How to virtualize radio infrastructure for multi-tenant operation? Share the same physical cells across automotive, enterprise, and video tenants, each with its own SLA.
4. How to keep fronthaul latency within tight budgets? HARQ¹ deadlines give you ~250 µs one-way between radio head and baseband processor (Rost et al. 2014; 3GPP 2022b).

---

5.2 Act 1: “It’s 1991. Europe Needs a Unified Digital Cellular Standard.”

Europe in 1991 had fragmented analog cellular systems — NMT in Scandinavia, TACS in Britain, C-Netz in Germany — that could not interoperate. ETSI’s GSM group delivered the first commercial digital cellular network that year: a standard spanning transmit frequency, modulation, multiplexing (TDMA on 200 kHz channels), mobility management, and billing (Dahlman et al. 2020; MacDonald 1979). The architecture was monolithic by necessity. Custom DSP silicon executed baseband processing; the Base Transceiver Station (BTS) housed radios, baseband, and control logic in one cabinet; the Mobile Switching Center (MSC) housed call routing, subscriber databases, and billing in another.

What the pioneers saw: A voice service. Fixed 13 kbps voice codec. Calls last minutes. Users walk or drive between cells at predictable speeds. Traffic is symmetric, continuous during the call, and bounded by the number of simultaneous calls per cell. The state per call is small: radio resources assigned, voice codec parameters, billing meter. The MSC can sit at the heart of a region and route every call.

What remained invisible from the pioneers’ vantage point: Packet data would come, then dominate. Voice would shrink to a minority of traffic. Users would stream video, play games, and run enterprise applications demanding different latency and reliability guarantees. The vertically integrated BTS + MSC would become the bottleneck: every new service required a vendor upgrade to the monolith.

5.2.1 The Solution: Monolithic BTS + MSC

GSM applied decision placement by centralizing everything — radio control at the BTS, session control at the BSC (Base Station Controller) that managed multiple BTSs, and call routing + subscriber authority at the MSC. State was bound to specific hardware: the Home Location Register (HLR) was an appliance; the Visitor Location Register (VLR) was an appliance; the MSC was an appliance. Interfaces (A-bis between BTS and BSC, A between BSC and MSC, MAP between MSC and HLR) were standardized but proprietary — vendor-specific protocol stacks running on vendor-specific hardware (Dahlman et al. 2020).

5.2.2 Invariant Analysis: GSM BTS + MSC (1991)

Invariant	GSM Answer (1991)	Gap?
State	Per-call circuit state in BTS + MSC	Hardware-bound; cannot migrate
Time	Circuit-switched, fixed allocation	Statistical multiplexing absent
Coordination	MSC orchestrates region	Single point of failure
Interface	Proprietary A-bis, A, MAP	Vendor lock-in

The State gap is that every subscriber’s registration is pinned to a specific MSC/HLR (Dahlman et al. 2020); a hardware failure requires physical replacement. The Time gap is that the circuit-switched model wastes bandwidth during silent portions of a call and cannot absorb bursty data traffic. The Coordination gap is that the MSC is a scaling bottleneck: adding capacity means adding MSCs, each with their own subscriber database that must synchronize. The Interface gap is that operators are locked to a single vendor per network — a Nokia BTS speaks only to a Nokia BSC.

5.2.3 Environment → Measurement → Belief

Layer	What GSM Has	What’s Missing
Environment	Users with voice calls, moving between cells	—
Measurement	Received signal strength per neighbor cell	Packet-level traffic awareness absent
Belief	“This call’s resources are reserved on this BTS”	Data-demand belief absent

The E→M gap is structurally filtered: the measurement signal (RSSI reports from handsets) answers voice-era questions (“which cell should serve this call?”) but says nothing about data demand. The belief model has no vocabulary for packets.

5.2.4 “The Gaps Didn’t Matter… Yet.”

Voice was the service. SMS was a happy accident riding unused control-channel capacity. Subscribers carried phones, not computers. The monolith matched the service. The 13 kbps codec matched the fixed 200 kHz TDMA slot. The MSC matched the call-routing model. For a decade, GSM worked.

Then packet data arrived — web browsing, email, messaging — and the circuit-switched assumption broke. A user checking email held a circuit for seconds of traffic + minutes of idle, consuming radio resources for silence. The monolith carried IP packets only with massive per-call overhead.

---

5.3 Act 2: “It’s 2000. Packet Data Demands a Parallel Path.”

3GPP released UMTS (Rel-99) in 1999–2000 as the successor to GSM (3GPP 2022c). Rather than rebuild the core, the designers bolted a packet-switched path alongside the circuit-switched core. The base station (now called NodeB) branched: voice traffic continued to the MSC via the Radio Network Controller (RNC); packet traffic flowed through a new path, SGSN (Serving GPRS Support Node) → GGSN (Gateway GPRS Support Node), onto the public Internet.

What the pioneers saw: Voice must keep working (revenue). Data is new and uncertain in volume. The safest architectural move is parallelism — preserve the circuit path intact; add a packet path next to it. Reuse as much of GSM as possible (HLR, authentication, billing).

What remained invisible: Data would overtake voice in volume within a decade. Operating two parallel cores (circuit + packet) became an OpEx burden. Every handoff would need to coordinate both paths. The parallel architecture was a waypoint toward full IP convergence.

5.3.1 The Solution: Bifurcated Core

UMTS applied disaggregation by separating the packet data path from the circuit voice path — two coexisting stacks sharing the RAN and the subscriber database. The NodeB handed uplink packets to the RNC (3GPP 2022c), which demultiplexed: voice bearers to the MSC, data bearers to the SGSN. The SGSN anchored the user’s packet session; the GGSN was the Internet gateway.

5.3.2 Invariant Analysis: UMTS Bifurcated Core (2000)

Invariant	UMTS Answer (2000)	Gap?
State	Circuit state + packet state in parallel	Duplicated subscriber context
Time	Voice: fixed circuit; data: best-effort packets	Two timing models coexist
Coordination	MSC (voice) and SGSN/GGSN (data) coexist	Handoff coordinates both paths
Interface	ATM on Iub and Iu-CS; IP on Iu-PS	Mixed transport

The State gap is duplicate subscriber context (3GPP 2022c): HLR authentication must be queried by both the MSC (for voice) and the SGSN (for data), doubling signaling. The Time gap is that voice and data paths run on different timing assumptions — circuit scheduling for voice, best-effort queueing for data. The Coordination gap is that mobility requires dual-path handoff: the MSC updates the VLR while the SGSN updates its packet context, and both must complete before the new cell can serve the user. The Interface gap is transport heterogeneity: ATM for voice, IP for data, on the same physical backhaul.

5.3.3 “The Gaps Didn’t Matter… Yet.”

Data volumes remained modest through the early 2000s. HSPA (Rel-5, 6) pushed packet throughput to megabits per second (Dahlman et al. 2020; 3GPP 2022c), but voice still drove revenue. The parallel architecture was tolerable. Then the iPhone shipped in 2007. Smartphone data consumption per subscriber grew 10× in two years. Operators found two parallel cores economically unsustainable. A unified all-IP architecture became unavoidable.

---

5.4 Act 3: “It’s 2008. All Traffic Must Become IP.”

3GPP Release 8 (2008) delivered LTE: an all-IP architecture that eliminated the circuit core (3GPP 2022a). Voice became VoIP (VoLTE). The base station became the eNodeB (eNB), integrating RNC functions for lower latency (3GPP 2022a). The core collapsed into four functions: MME (Mobility Management Entity, control plane), S-GW (Serving Gateway, user plane anchor), P-GW (PDN Gateway, Internet gateway), HSS (Home Subscriber Server, subscription database) (3GPP 2022a). Critically, LTE introduced the X2 interface — direct eNB-to-eNB signaling for handoff coordination that bypassed the core entirely (3GPP 2022a).

What the pioneers saw: IP is the universal transport. Voice can be a packet service. Smartphones need low-latency handoff to keep TCP connections alive. If eNBs can coordinate directly (X2), the core is bypassed for handoff mediation — handoff latency drops from seconds to under 150 ms.

What remained invisible: URLLC (ultra-reliable low-latency), mMTC (massive IoT), and eMBB (extreme broadband) would demand three radically different service profiles on the same infrastructure (Andrews et al. 2014; NGMN Alliance 2015). One-size-fits-all scheduling was insufficient. The flat EPC with monolithic S-GW and P-GW would become a scaling bottleneck.

5.4.1 The Solution: eNB + Flat EPC + X2

LTE applied disaggregation by splitting control and user planes across specialized core functions (3GPP 2022a): MME (control-only, never sees user packets), S-GW (user-plane anchor inside the PLMN), P-GW (user-plane gateway to the Internet). The eNodeB absorbed RNC functions, reducing hops in the user path. Decision placement moved handoff coordination to the edge (X2) for latency while keeping attachment and paging centralized at the MME (3GPP 2022a).

5.4.2 Invariant Analysis: LTE eNB + EPC (2008)

Invariant	LTE Answer (2008)	Gap?
State	EPS Bearer per user, anchored at S-GW	Service-agnostic bearer model
Time	TTI = 1 ms; X2 handoff 50-150 ms	Can’t meet URLLC sub-ms
Coordination	MME orchestrates; X2 peer-peer for HO	Monolithic MME at scale
Interface	S1-AP, X2-AP, GTP-U (proprietary telco)	Not cloud-native

The State gap is that all EPS bearers share the same QCI (QoS Class Identifier) ontology (3GPP 2022a) — fine for voice + web, coarse for distinguishing URLLC from video. The Time gap is hard: TTI = 1 ms and handoff = 50-150 ms both assume stable physical infrastructure; URLLC will demand sub-ms deterministic scheduling that eNBs cannot deliver without architectural change. The Coordination gap is that the MME becomes a single-point-of-scaling (every attach/handoff signals it). The Interface gap is that GTP-U, S1-AP, and X2-AP are telco-specific protocols, orthogonal to the cloud-native HTTP/REST world.

5.4.3 Environment → Measurement → Belief After LTE

Layer	What LTE Has	What’s Missing
Environment	Mixed smartphone traffic across mobile users	—
Measurement	CQI (Channel Quality Indicator), RSRP, RSRQ per UE; handover measurements	No per-service SLA awareness
Belief	“This bearer has QCI=9 (best-effort)”	No belief about slice or tenant

The E→M gap is structurally filtered by the QCI ontology: everything maps into 9 pre-defined classes. For URLLC — 1 ms latency, 10⁻⁵ packet loss — existing QCIs are inadequate, and the bearer model lacks vocabulary to encode “this user is part of the automotive slice with guaranteed 5 ms end-to-end budget.”

5.4.4 “The Gaps Didn’t Matter… Yet.”

LTE was the golden era of mobile broadband. Smartphones multiplied; users streamed video; X2 handoff kept TCP alive at highway speeds. Then operator economics shifted. Hardware costs rose faster than subscription ARPU. Densification (more cells) became unavoidable, and each new cell required a new eNodeB with full baseband processing (3GPP 2022a). The per-cell BBU was underutilized 80% of the time but paid for at peak (China Mobile Research Institute 2011). Centralizing baseband into a shared pool — Cloud RAN — promised 20–30% hardware savings (Checko et al. 2015; China Mobile Research Institute 2011).

---

5.5 Act 4: “It’s 2018. 5G Demands Split, Scalable, Service-Aware Infrastructure.”

3GPP Release 15 (2018) delivered 5G NR and the 5G Core (3GPP 2022b, 2022d). Two architectural moves defined the generation. First, the base station was split: gNodeB (gNB) disaggregated into CU (Central Unit) + DU (Distributed Unit) + RU (Radio Unit) connected by standardized interfaces F1, E1, and Open Fronthaul. Second, the core was atomized into a Service-Based Architecture² (SBA): stateless Network Functions³ (AMF (Access and Mobility Management Function), SMF (Session Management Function), UPF (User Plane Function), UDM (Unified Data Management), PCF (Policy Control Function), NRF, AUSF (Authentication Server Function), NSSF (Network Slice Selection Function)) communicating over HTTP/2 REST APIs (3GPP 2022d).

What the pioneers saw: URLLC needs DU close to the antenna (sub-ms RTT). eMBB benefits from CU pooling (centralized scheduling across many cells). mMTC benefits from stateless NFs that scale horizontally in the cloud. One physical network must carry three service profiles. The only way to reconcile these is to disaggregate along the axes that matter: latency (CU/DU/RU split), state (control vs user plane), and function (decompose core into microservices).

What remained invisible: The service models running between RIC and CU/DU would become the new bottleneck. Vendor interoperability at open interfaces (especially Open Fronthaul) would take years to mature. The fronthaul latency budget⁴ (~250 µs one-way) would limit DU placement to within 20 km of the RU (Rost et al. 2014; 3GPP 2022b).

5.5.1 The Solution: gNB Split + 5GC SBA

5G applied disaggregation at two layers simultaneously. At the RAN: CU (RRC (Radio Resource Control) + PDCP (Packet Data Convergence Protocol) + SDAP, 10-100 ms timescale) + DU (RLC (Radio Link Control) + MAC + High-PHY, ms timescale) + RU (Low-PHY + RF, µs timescale). The split points (3GPP Option 2 at PDCP/RLC for CU-DU; Option 7-2x⁵ inside PHY for DU-RU) emerged from the fronthaul bandwidth-latency tradeoff (Rost et al. 2014; Dahlman et al. 2020; 3GPP 2022b). At the core: the monolithic EPC shattered into ~15 network functions, each independently scalable. Closed-loop reasoning entered the infrastructure itself: the SBF (Service-Based Framework) uses NRF for service discovery, enabling NFs to find and invoke each other dynamically (3GPP 2022d).

5.5.2 Invariant Analysis: 5G gNB + 5GC (2018)

Invariant	5G Answer (2018)	Gap?
State	Per-UE context split across CU/DU/RU + NFs	State sync across splits
Time	Multi-timescale: µs (RU) → ms (DU) → 100ms (CU)	RIC timescale not yet defined
Coordination	Hierarchical CU→DU→RU + NRF service discovery	Cross-NF state consistency
Interface	F1AP, E1AP, Xn, HTTP/2 SBI, Open Fronthaul	Multi-vendor integration cost

The State gap is that the per-UE context now lives in four places (RU HARQ buffers, DU MAC/scheduler, CU RRC/PDCP, AMF registration) (3GPP 2022b) and mobility requires consistent migration of all four. The Time gap is that the emerging RIC layer (control loops on top of gNB) does not yet have a standardized timescale — does it run at 10 ms, 100 ms, 1 s? The Coordination gap is cross-NF consistency: if SMF-A writes session state and SMF-B reads it milliseconds later, eventual consistency in the session database can corrupt the session. The Interface gap is multi-vendor cost: Open Fronthaul defines the interface, but vendor A’s RU and vendor B’s DU still require integration testing.

5.5.3 Environment → Measurement → Belief After 5G NG-RAN

Layer	What 5G NG-RAN Has	What’s Missing
Environment	Heterogeneous services across one RAN	—
Measurement	CU telemetry, DU scheduler stats, per-NF metrics	Standardized ML feature pipeline absent
Belief	Per-NF operational state	Global optimization objective absent

The E→M gap is structurally filtered: telemetry exists at each NF but remains unstandardized as a coherent feature set for ML. Each vendor exposes its own counters. Cross-vendor optimization requires a common service model that remains absent.

5.5.4 “The Gaps Didn’t Matter… Yet.”

Early 5G deployments (2019–2021) were eMBB-focused: faster speeds on smartphones. The split architecture delivered capacity. Service heterogeneity did not materialize at scale. Then URLLC pilots (private 5G factories) and network slicing SLA contracts began exposing the ML-feature gap. Operators needed programmable control loops — adaptive closed loops beyond the static QoS parameters of 3GPP that could respond to real-time conditions. Closed vendor RANs blocked this path. The binding constraint inverted: from vendor control of the stack to operator control of the logic.

---

5.6 Act 5: “It’s 2020. Operators Demand Open, Programmable RAN Control.”

The O-RAN Alliance formed in 2018 out of two predecessor groups (xRAN Forum + C-RAN Alliance). By 2020 it had published the first architecture description (O-RAN Alliance 2022): open interfaces built on 3GPP’s CU/DU/RU split, plus a RIC (RAN Intelligent Controller) layer for ML-driven optimization (Polese et al. 2023; Bonati et al. 2020). Two RICs were defined (O-RAN Alliance 2022; Polese et al. 2023): Near-RT RIC (10 ms – 1 s loops, running xApps⁶, connected to CU/DU via E2⁷) and Non-RT RIC (>1 s loops, running rApps⁸, connected to Near-RT RIC via A1⁹, hosted inside the SMO). O1¹⁰ connects the SMO to all network elements for management. Open Fronthaul standardized the 7-2x split for multi-vendor RU + DU interoperability.

What the pioneers saw (the Alliance operators): Closed RAN is a cost and velocity problem. Every optimization — better scheduling, smarter beamforming, slice-aware QoS — must wait for the incumbent vendor’s release cycle. Opening E2 lets operators deploy third-party xApps (O-RAN Alliance 2022). Opening A1 lets them push declarative policy from SMO (O-RAN Alliance 2022). Opening fronthaul lets them mix RU and DU vendors (Polese et al. 2023).

What remained invisible: Running multiple xApps with conflicting objectives is an unsolved control problem. E2 telemetry at 1000 cells × fine granularity is a data-plane scaling problem. ML model lifecycle (training data, drift detection, rollback) is a MLOps problem imported into telco infrastructure. The binding constraint persists; it moves.

5.6.1 The Solution: Open Interfaces + RIC

O-RAN applied disaggregation along a new axis: control logic separated from the data path. The gNB continues to handle fast-path scheduling; the RIC observes telemetry, runs ML, and injects policy via standardized service models. Applied closed-loop reasoning at three nested timescales (Polese et al. 2023; Bonati et al. 2020): MAC scheduler (TTI, ~1 ms) → Near-RT RIC (10 ms – 1 s) → Non-RT RIC (>1 s). Each loop’s bandwidth is deliberately slower than the one below to avoid oscillation. Applied decision placement on an explicit continuum: the fastest decisions stay at the DU (because latency is physical), while strategic decisions migrate to the Non-RT RIC (because global view is needed).

5.6.2 Invariant Analysis: O-RAN with RIC (2020+)

Invariant	O-RAN Answer (2020+)	Gap?
State	Near-RT RIC: aggregated UE/cell KPIs; Non-RT RIC: training data	xApp state consistency
Time	Three loops: MAC TTI, Near-RT (10ms-1s), Non-RT (>1s)	Loop interaction not formalized
Coordination	Hierarchical control via A1/E2	xApp conflict resolution unsolved
Interface	E2, A1, O1, Open Fronthaul	Multi-vendor integration tax

The State gap is xApp state consistency (Polese et al. 2023): multiple xApps subscribing to overlapping E2 reports can derive conflicting beliefs. The Time gap is that loop-interaction dynamics (does a fast xApp destabilize a slower rApp?) are not formally specified. The Coordination gap is xApp conflict resolution: if xApp-A wants to hand off user U to cell X and xApp-B wants to keep U on cell Y, which wins? The Interface gap is the integration tax: each new open interface multiplies the test matrix (vendor combinations).

5.6.3 Environment → Measurement → Belief After O-RAN

Layer	What O-RAN Has	What’s Missing
Environment	Multi-vendor RAN with heterogeneous traffic	—
Measurement	Standardized E2SM-KPM reports, ML-ready	Cross-xApp consistency
Belief	Each xApp holds its own belief model	Cross-xApp arbiter absent

The E→M gap is structurally filtered but by a new filter: the E2 service model (E2SM-KPM) defines what can be reported. Vendors must implement the service model to expose data. Expanding the service model requires Alliance consensus.

5.6.4 “The Gaps Didn’t Matter… Yet.”

In single-vendor pilots, xApp conflicts are rare (the vendor ships one coherent set). As the xApp ecosystem grows, the conflict problem will sharpen. The xApp conflict question is today what BGP policy oscillation was in the late 1990s: a distributed control problem whose pathologies emerge only at scale.

---

5.7 Act 6: “It’s 2020+. One Infrastructure Must Serve Many Tenants.”

Parallel with O-RAN, 3GPP Release 15–17 standardized network slicing (Foukas et al. 2017; 3GPP 2022d): end-to-end virtual networks with independent SLAs over shared physical infrastructure. A slice is identified by S-NSSAI (Single Network Slice Selection Assistance Information — a composite of Slice/Service Type + Slice Differentiator)¹¹ (3GPP 2022d); the NSSF routes a UE’s attach to the right AMF instance for its slice (3GPP 2022d); slice-specific SMF + UPF chains enforce slice QoS (Foukas et al. 2017); RAN-side slicing maps slice IDs to scheduling priorities and dedicated resource pools.

What the pioneers saw: One 5G network must serve mobile broadband subscribers + autonomous vehicles + IoT sensors + private enterprise + public safety, each with different SLA requirements. Building parallel networks is prohibitively costly. Slicing amortizes capital across tenants.

What remained invisible: The isolation/multiplexing tradeoff has no free lunch. Hard isolation (dedicated resources per slice) destroys multiplexing gain. Soft isolation (shared resources, priority arbitration) weakens SLA guarantees under load. Mobility across slices complicates session continuity when a UE moves between cells that implement different slice configurations.

5.7.1 The Solution: End-to-End Logical Networks

Network slicing applied disaggregation to the network itself — one physical infrastructure, N logical networks, each with independent NFs, state, and policy. Applied closed-loop reasoning at the orchestration layer: the NSMF (Network Slice Management Function) monitors per-slice SLA compliance and reallocates resources when drift is detected. Applied decision placement by splitting admission (centralized orchestrator, global view) from enforcement (distributed NFs, per-packet).

5.7.2 Invariant Analysis: 5G Network Slicing (2020+)

Invariant	Slicing Answer (2020+)	Gap?
State	Per-slice resource pools + S-NSSAI in UE context	Cross-slice mobility state
Time	Multi-timescale: SLA (days), reallocation (s), enforcement (ms)	Slice dimensioning under burstiness
Coordination	NSMF central admission + per-NF enforcement	Multi-domain (RAN+transport+core) orchestration
Interface	S-NSSAI, NSSF, slice management APIs (TS 28.530)	Cross-operator slice federation immature

The State gap is cross-slice mobility: a UE moving between cells with different slice configurations must renegotiate slice membership mid-session. The Time gap is slice dimensioning: traffic burstiness pushes a slice over its budget before the orchestrator reacts. The Coordination gap is orchestrating across RAN + transport + core — each domain has its own management stack. The Interface gap is cross-operator federation: slicing today is intra-operator; inter-operator slice SLAs (for roaming URLLC, for instance) remain immature.

5.7.3 The Tension: Isolation vs Multiplexing Gain

Hard isolation: every slice gets dedicated PRBs, UPF instances, processing. SLAs are guaranteed. Multiplexing gain is zero. Cost is high. Soft isolation: slices share resources with priority arbitration. Multiplexing gain is high. SLAs degrade under contention. Real deployments blend both: a guaranteed minimum allocation (hard) + best-effort oversubscription (soft).

5.7.4 Environment → Measurement → Belief After Slicing

Layer	What Slicing Has	What’s Missing
Environment	Multi-tenant heterogeneous demand	—
Measurement	Per-slice KPIs, SLA compliance metrics	Predictive demand models
Belief	“Slice X is at Y% of its budget”	Anticipatory reallocation absent

The E→M gap is accidentally noisy — per-slice telemetry exists but demand prediction (to reallocate before SLA violation) is an open ML problem.

---

5.8 The Grand Arc: From Monolithic BTS to Programmable Platform

5.8.1 The Evolving Anchor

Era	Derived Constraint	Architectural Answer	Binding Scaling Limit
GSM (1991)	Voice-only service	Monolithic BTS + MSC	MSC capacity
UMTS (2000)	Packet data bolted on	Bifurcated core (SGSN/GGSN parallel to MSC)	Dual-path OpEx
LTE (2008)	All-IP unification	eNB + flat EPC + X2	Monolithic S/P-GW
5G (2018)	Service heterogeneity	CU/DU/RU split + 5GC SBA	Fronthaul latency + NF state consistency
O-RAN (2020+)	Vendor lock-in	Open interfaces + RIC	xApp conflict + integration tax
Slicing (2020+)	Multi-tenant sharing	End-to-end logical networks	Isolation/multiplexing tradeoff

The binding constraint at the physical layer never changed: licensed spectrum, exclusive operator ownership, mobility across cells. What changed was the derived constraint at higher layers — what service must be delivered, what economics must be met, what tenants must share.

5.8.2 Three Design Principles Applied Across the Arc

Disaggregation is the through-line of the entire chapter. Each generation separated what the previous generation had coupled. GSM’s monolith (Dahlman et al. 2020) became UMTS’s bifurcated core (3GPP 2022c), which became LTE’s flat EPC (3GPP 2022a), which became 5G’s CU/DU/RU + SBA (3GPP 2022b, 2022d), which became O-RAN’s control-plane/data-plane split (O-RAN Alliance 2022), which became slicing’s per-tenant logical networks (Foukas et al. 2017). Every separation created a new interface; every interface became a coordination point.

Closed-loop reasoning became explicit infrastructure only in 5G. GSM had implicit loops (power control, handoff). LTE added eICIC and CoMP at ms timescales. 5G introduced the RIC as a first-class loop (3GPP 2022b). O-RAN formalized a three-loop hierarchy (MAC TTI → Near-RT → Non-RT) (O-RAN Alliance 2022; Polese et al. 2023). Each loop’s bandwidth is calibrated to sit below the faster loop below and above the slower loop above — the classic control-theoretic cascade.

Decision placement traced a continuum from “all decisions at the MSC” (GSM) to “fast at DU, medium at Near-RT RIC, strategic at Non-RT RIC” (O-RAN). The invariant rule: fast decisions stay close to the radio because latency is physical; slow decisions migrate to centralized orchestrators because global view is needed.

5.8.3 The Dependency Chain

flowchart TD
    A[Licensed spectrum + operator monopoly] -->|locks| B[Centralized per-cell scheduling]
    B -->|plus mobility| C[Cross-cell coordination]
    C -->|GSM answer| D[Monolithic BTS + MSC]
    D -->|data demand breaks| E[UMTS bifurcated core]
    E -->|OpEx breaks| F[LTE all-IP + X2]
    F -->|service heterogeneity breaks| G[5G CU/DU/RU + SBA]
    G -->|vendor lock breaks| H[O-RAN open interfaces + RIC]
    G -->|multi-tenant demand| I[Network slicing]

    style A fill:#cce5ff,stroke:#004085
    style D fill:#f8d7da,stroke:#721c24
    style E fill:#f8d7da,stroke:#721c24
    style F fill:#f8d7da,stroke:#721c24
    style G fill:#d4edda,stroke:#155724
    style H fill:#d4edda,stroke:#155724
    style I fill:#d4edda,stroke:#155724

5.8.4 Pioneer Diagnosis Table

Year	Pioneer	Invariant	Diagnosis	Contribution
1991	ETSI GSM	State	Circuit state ties to hardware	Monolithic BTS+MSC works for voice (Dahlman et al. 2020)
2000	3GPP UMTS	Interface	Circuit core can’t carry IP	Parallel packet path (SGSN/GGSN) (3GPP 2022c)
2008	3GPP LTE	Coordination	MSC-mediated handoff too slow	X2 direct eNB-eNB coordination (3GPP 2022a)
2010	China Mobile	State	Per-cell BBU underutilized	C-RAN centralized BBU pool (China Mobile Research Institute 2011)
2017	3GPP CUPS	Interface	S/P-GW coupling prevents edge UPF	Control/user plane separation12 (Dahlman et al. 2020)
2018	3GPP NG-RAN	Coordination	Functions need different timescales	CU/DU/RU split with F1/E1 (3GPP 2022b)
2018	3GPP 5GC	Interface	Telco protocols not cloud-native	Service-Based Architecture (HTTP/2) (3GPP 2022d)
2018	O-RAN Alliance	Interface	Closed RAN prevents innovation	Open interfaces + RIC (O-RAN Alliance 2022)
2020+	3GPP slicing	State	Shared infrastructure, heterogeneous tenants	S-NSSAI + NSSF + per-slice NFs (Foukas et al. 2017; 3GPP 2022d)

5.8.5 Innovation Timeline

flowchart TD
    subgraph sg1["Monolithic"]
        A1["1991 — GSM BTS + MSC (voice)"]
        A2["2000 — UMTS NodeB + SGSN/GGSN (packet bolted on)"]
        A1 --> A2
    end
    subgraph sg2["All-IP"]
        B1["2008 — LTE eNB + EPC + X2"]
        B2["2010 — C-RAN BBU pool"]
        B1 --> B2
    end
    subgraph sg3["Disaggregated"]
        C1["2017 — CUPS (S-GW, P-GW split)"]
        C2["2018 — 5G gNB CU/DU/RU + 5GC SBA"]
        C1 --> C2
    end
    subgraph sg4["Programmable"]
        D1["2018 — O-RAN Alliance founded"]
        D2["2020 — Open Fronthaul + Near-RT RIC + Non-RT RIC"]
        D3["2020 — Network slicing (S-NSSAI, NSSF)"]
        D1 --> D2 --> D3
    end
    sg1 --> sg2 --> sg3 --> sg4

Cellular Infrastructure Evolution

---

5.9 Generative Exercises

TipExercise 1: Private 5G on a Factory Floor

A manufacturer deploys private 5G across a factory — 20 cells, 1000 devices, single administrative domain, no roaming. Predict how the invariant answers should shift:

• State: Can state centralize (one AMF, one SMF instance) since there are no roaming partners?
• Time: Can handoff tighten below 50 ms since all cells are within one LAN?
• Coordination: Can scheduling centralize across all cells (joint PRB allocation) since latency between cells is sub-ms?
• Interface: What must remain 3GPP-compliant (for UE compatibility) vs what can go proprietary?

Use the disaggregation principle to argue which NFs should remain separate even in a small deployment, and which can collapse.

TipExercise 2: URLLC Slice for Autonomous Vehicles

An automotive tenant demands an URLLC slice: 10 ms end-to-end latency, 99.999% reliability, vehicles moving at 120 km/h. Predict the architectural constraints:

• Where must the UPF live? (Answer through the latency budget.)
• What fronthaul split option is compatible with the latency budget? (Argue using CPRI (Common Public Radio Interface — the legacy fronthaul protocol carrying raw I/Q samples) vs eCPRI (enhanced CPRI — the newer protocol carrying partially processed data at lower bandwidth) bandwidth.)
• How frequently can handoff occur at 120 km/h? (Estimate cell residency times.)
• What does hard isolation cost in PRB utilization during low-vehicle periods?

Then argue: does URLLC force dedicated infrastructure, or can it share with eMBB via preemption?

TipExercise 3: xApp Conflict Resolution

Two xApps run on the same Near-RT RIC. xApp-A (throughput optimizer) wants to hand UE U to cell X (higher MCS). xApp-B (energy optimizer) wants to keep U on cell Y (keeps cell X in low-power state). Design a conflict resolution policy:

• Priority-based: rank xApps statically. What could go wrong?
• Policy-based: Non-RT RIC rApp arbitrates via A1. What is the latency cost?
• Objective-based: combine objectives into one score. What assumption does this require?

Argue which approach matches the BGP policy oscillation lesson from Chapter 9.

---

5.10 References

• 3GPP TS 38.300, “NR; NR and NG-RAN Overall Description; Stage 2” (3GPP 2022b)
• 3GPP TS 23.501, “System Architecture for the 5G System (5GS); Stage 2” (3GPP 2022d)
• O-RAN Alliance, “O-RAN Architecture Description v07.00” (O-RAN Alliance 2022)
• Foukas et al., “Network Slicing in 5G: Survey and Challenges” (Foukas et al. 2017)
• Checko et al., “Cloud RAN for Mobile Networks — A Technology Overview” (Checko et al. 2015)
• Polese et al., “Understanding O-RAN: Architecture, Interfaces, Algorithms, Security, and Research Challenges” (Polese et al. 2023)
• NGMN Alliance, “5G White Paper” (NGMN Alliance 2015)
• Bonati et al., “Open, programmable, and virtualized 5G networks” (Bonati et al. 2020)
• Dahlman, Parkvall, Sköld, “5G NR: The Next Generation Wireless Access Technology” (Dahlman et al. 2020)
• Rost et al., “Cloud Technologies for Flexible 5G Radio Access Networks” (Rost et al. 2014)
• China Mobile Research Institute, “C-RAN: The Road Towards Green RAN” (China Mobile Research Institute 2011)
• Andrews et al., “What Will 5G Be?” (Andrews et al. 2014)

---

This chapter is part of “A First-Principles Approach to Networked Systems” by Arpit Gupta, UC Santa Barbara, licensed under CC BY-NC-SA 4.0.

3GPP. 2022a. E-UTRA and E-UTRAN Overall Description; Stage 2. TS 36.300. 3rd Generation Partnership Project.

3GPP. 2022b. NR; NR and NG-RAN Overall Description; Stage 2. TS 38.300. 3rd Generation Partnership Project.

3GPP. 2022c. Radio Interface Protocol Architecture. TS 25.301. 3rd Generation Partnership Project.

3GPP. 2022d. System Architecture for the 5G System (5GS); Stage 2. TS 23.501. 3rd Generation Partnership Project.

Andrews, Jeffrey G., Stefano Buzzi, Wan Choi, et al. 2014. “What Will 5G Be?” IEEE Journal on Selected Areas in Communications 32 (6): 1065–82.

Bonati, Leonardo, Michele Polese, Salvatore D’Oro, Stefano Basagni, and Tommaso Melodia. 2020. “Open, Programmable, and Virtualized 5G Networks: State-of-the-Art and the Road Ahead.” Computer Networks 182: 107516.

Checko, Aleksandra, Henrik L. Christiansen, Ying Yan, et al. 2015. “Cloud RAN for Mobile Networks — a Technology Overview.” IEEE Communications Surveys & Tutorials 17 (1): 405–26.

China Mobile Research Institute. 2011. C-RAN: The Road Towards Green RAN. China Mobile.

Dahlman, Erik, Stefan Parkvall, and Johan Sköld. 2020. 5G NR: The Next Generation Wireless Access Technology. 2nd ed. Academic Press.

Foukas, Xenofon, Georgios Patounas, Ahmed Elmokashfi, and Mahesh K. Marina. 2017. “Network Slicing in 5G: Survey and Challenges.” IEEE Communications Magazine 55 (5): 94–100.

MacDonald, V. H. 1979. “The Cellular Concept.” The Bell System Technical Journal 58 (1): 15–41.

NGMN Alliance. 2015. 5G White Paper. Next Generation Mobile Networks Alliance.

O-RAN Alliance. 2022. O-RAN Architecture Description v07.00. O-RAN.WG1.O-RAN-Architecture-Description-v07.00. O-RAN Alliance.

Polese, Michele, Leonardo Bonati, Salvatore D’Oro, Stefano Basagni, and Tommaso Melodia. 2023. “Understanding O-RAN: Architecture, Interfaces, Algorithms, Security, and Research Challenges.” IEEE Communications Surveys & Tutorials 25 (2): 1376–411.

Rost, Peter, Carlos J. Bernardos, Antonio De Domenico, et al. 2014. “Cloud Technologies for Flexible 5G Radio Access Networks.” IEEE Communications Magazine 52 (5): 68–76.

---

1. HARQ (Hybrid Automatic Repeat Request) combines forward error correction with retransmission. The receiver attempts to decode the received data; if decoding fails, it requests a retransmission and combines the original and retransmitted copies for a stronger signal. HARQ timing is strict: the UE expects an ACK/NACK within a fixed number of slots, creating a hard deadline for the fronthaul path.

2. A service-based architecture (SBA) replaces point-to-point interfaces between fixed network elements with a publish/subscribe model: each network function registers its services with a central repository (NRF) and discovers other functions dynamically via HTTP/2 REST APIs. This is the telco equivalent of microservices — each function is independently deployable, scalable, and replaceable.

3. A Network Function (NF) is a software-defined processing element that implements one specific piece of core network logic — authentication, session management, policy enforcement, etc. In pre-5G architectures, these were hardware appliances; in 5G, they are containerized microservices.

4. The ~250 µs one-way fronthaul budget is dictated by HARQ (Hybrid Automatic Repeat Request) timing: the UE expects an ACK/NACK within a fixed number of slots after transmitting. In 5G NR with 30 kHz subcarrier spacing, the HARQ round-trip is ~4 ms, leaving roughly 250 µs for the one-way fronthaul path between RU and DU. At the speed of light in fiber (~200 m/µs), this limits DU placement to within ~20 km of the RU.

5. Option 7-2x places the split inside the physical layer: the RU performs low-PHY functions (FFT/IFFT, cyclic prefix, beamforming) while the DU performs high-PHY functions (channel coding, modulation, HARQ processing). This split reduces fronthaul bandwidth from raw I/Q samples (Option 8, CPRI) to partially processed frequency-domain symbols, cutting bandwidth requirements by roughly 10x while keeping latency within the ~250 µs fronthaul budget.

6. An xApp is a third-party application that runs on the Near-RT RIC and implements a specific RAN optimization — e.g., handover optimization, interference management, or slice-aware scheduling. xApps consume E2 telemetry and inject control actions back via E2.

7. E2 is the O-RAN interface connecting the Near-RT RIC to the CU and DU. It carries both telemetry (E2 Service Model reports from CU/DU to RIC) and control actions (RIC to CU/DU). The E2 Service Model (E2SM) defines what data can be reported and what actions can be taken.

8. An rApp is an application running on the Non-RT RIC that implements strategic, slower-timescale optimization — e.g., ML model training, policy generation, or network-wide analytics. rApps push policy to the Near-RT RIC via A1.

9. A1 is the O-RAN interface from the Non-RT RIC to the Near-RT RIC. It carries declarative policy (e.g., “prioritize URLLC slice”) and ML model updates, but not per-UE control actions.

10. O1 is the O-RAN interface for management and orchestration, connecting the SMO (Service Management and Orchestration) to the O-RAN network elements (CU, DU, RU, Near-RT RIC). It handles configuration, performance management, and fault management.

11. S-NSSAI is the identifier a UE presents during registration to request a specific network slice. The Slice/Service Type (SST) is a standardized 8-bit value (e.g., SST=1 for eMBB, SST=2 for URLLC, SST=3 for mMTC). The Slice Differentiator (SD) is an optional 24-bit value that distinguishes among slices of the same type — e.g., two different enterprise URLLC slices.

12. Control-user plane separation (CUPS) decouples signaling (session setup, policy, mobility) from data forwarding (packet routing). This allows the control plane to remain centralized in the cloud while user-plane functions (UPFs) are placed at the network edge, closer to users, reducing latency for data traffic without fragmenting control logic.