A First-Principles Approach to Networked Systems
Preface
Why First Principles
An LLM can recite how TCP’s three-way handshake works. It can describe AIMD. It can list the fields in an 802.11 frame header. Mechanism-level recall is automated. The question every networking student and engineer now faces: what knowledge is durable when the machine already knows the facts?
The answer: first-principles reasoning — the ability to identify the binding constraint of an unfamiliar system, trace how one design choice forces the rest, and predict what breaks when the environment changes. An LLM that recites TCP’s congestion control lacks the judgment to decide whether that mechanism belongs in a system with 10 µs RTTs, one with 22-minute Mars links, or one running inside an agentic orchestrator. A human trained in first-principles reasoning provides that judgment. That human, working with the LLM, produces results that surpass either working alone.
Trained human + AI > AI alone > untrained human alone.
This book produces the trained human. It teaches networking as systems design under constraint, organized around four structural questions that every networked system answers:
- State — What does the system know about its environment, and where does that knowledge fail?
- Time — When do things happen? How fast must decisions be?
- Coordination — Who decides? One entity or many?
- Interface — How do components interact? What crosses the boundary, and what stays hidden?
These four invariants define the design space. The answers differ radically — ALOHA has zero state, BGP has policy-encoded state, massive MIMO has high-dimensional channel state — yet the questions remain constant across every system and every era. A student who internalizes them can analyze any system, from 1970 packet radio to 2026 agentic AI, with the same vocabulary. That student becomes the expert who knows what to ask the LLM, how to structure the problem, and when the LLM’s output is wrong. First-principles reasoning is the lever that makes AI collaboration productive.
Without that lever, AI operates unguided — generating plausible-sounding configurations that violate constraints invisible to the model, proposing architectures that conflate invariants any practitioner would separate, diagnosing failures with confident explanations that miss the binding constraint entirely. The networking systems of the next decade will be built by humans and AI working together. This book equips the human side of that partnership.
The Framework
The analytical framework has three components:
Four invariants define what must be answered: State, Time, Coordination, Interface. Every system answers them, regardless of whether the designer addresses them explicitly. Chapter 1 introduces them with worked examples (TCP, DNS, DHCP, WiFi) that recur throughout the book.
Three design principles describe recurring strategies for constructing good answers:
- Disaggregation — separate concerns so each can evolve independently.
- Closed-loop reasoning — feedback loops stabilize adaptive systems. The Environment → Measurement → Belief decomposition diagnoses where loops fail.
- Decision placement — where control authority resides, from fully distributed to fully centralized.
The anchored dependency graph provides the analytical method. Every system has a binding constraint — the hardest-to-change invariant answer, inherited from the environment or a lower layer. The binding constraint locks one invariant, which forces the others in cascade.
TCP’s chain illustrates. Interface (IP datagrams, inherited) → Coordination (distributed, because IP provides only best-effort delivery) → State (inferred from local ACK signals) → Time (estimated via Jacobson’s RTT algorithm). Three researchers discovered three links in this chain independently, across three years — yet they converged on the same decomposition. Tracing such chains is how the book reconstructs each system’s design logic.
How the Book Is Organized
The book follows the Internet’s layered architecture bottom-up, from the physical medium to the application layer and beyond. Each chapter applies the full framework to a single system or system class: pioneer-driven narrative, invariant analysis, E-M-B decomposition, and dependency chain.
Part I: Foundations
Chapter 1 introduces the four invariants, three design principles, and the anchored dependency graph through motivating examples.
Part II: The Link Layer — Sharing a Physical Medium
The binding constraint inherited from the physical layer: electromagnetic spectrum is finite, shared, and noisy.
- Chapter 2 — Wired Link Layer: Ethernet CSMA/CD → STP → switches → fat-tree fabrics → VXLAN/EVPN
- Chapter 3 — Wireless Link Layer: ALOHA → CSMA/CA → 802.11ax OFDMA → cellular CDMA → massive MIMO
- Chapter 4 — Wireless Infrastructure: Monolithic BTS → LTE eNodeB → 5G CU/DU/RU → O-RAN → network slicing
Part III: The Network Layer — Forwarding, Routing, and Queue Management
The IP datagram is the narrow waist. Everything below delivers it; everything above uses it.
- Chapter 5 — Addressing, Naming, and Discovery: IP → DNS → CIDR → DHCP/ARP → NAT → IPv6
- Chapter 6 — Routing, Switching, and Programmable Networks: Distance-vector → link-state → BGP → MPLS → SDN → P4
- Chapter 7 — Queue Management: FIFO → fair queueing → RED → bufferbloat → CoDel → FQ-CoDel
Part IV: The Transport Layer — End-to-End Delivery
The IP interface forces transport to solve reliability and congestion control from endpoints only.
- Chapter 8 — Transport and Congestion Control: Cerf/Kahn → Jacobson → CUBIC → BBR → DCTCP → QUIC
- Chapter 9 — System Composition: Transport × queue coupling, L4S, PowerBoost, QoS vs QoE
Part V: The Application Layer — Services on the Stack
Applications inherit the transport interface and solve their own invariant problems.
- Chapter 10 — Application Protocols and Content Delivery: HTTP evolution → CDN → QUIC/HTTP3 → edge compute
- Chapter 11 — Multimedia Applications: RTP → DASH/HLS → MPC → WebRTC → low-latency streaming
Part VI: Cross-Cutting Concerns
- Chapter 12 — Measurement, Management, and the Research Frontier: BPF → streaming telemetry → INT → Sonata → verification
Part VII: Agentic Systems (Capstone)
- Chapter 13 — Agentic Systems: Prompt → CoT → ReAct → multi-agent → MCP thin waist → persistent memory. Closes the book’s arc: the four invariants apply to the newest system class — agents built on the infrastructure the preceding chapters describe.
How to Use This Book
Each chapter opens with the binding constraint that shapes the system under study. It walks through pioneer-driven Acts — chronological, each with invariant analysis and E-M-B tables — and closes with a Grand Arc revealing cross-generational patterns. Interactive visualizations (in the HTML edition) let you manipulate parameters and observe system dynamics directly.
The book serves upper-division CS undergraduates and beginning graduate students who have completed an introductory networking course. It assumes familiarity with TCP/IP, basic queuing, and protocol layering — but assumes nothing about how to reason about these systems. That reasoning is what it teaches.
The HTML edition includes interactive Observable JS visualizations that let you explore system dynamics by manipulating parameters directly. Look for the interactive figures throughout the text.
This book is under active development alongside CS 176C at UC Santa Barbara, Spring 2026.
Arpit Gupta, UC Santa Barbara Spring 2026
This book is licensed under CC BY-NC-SA 4.0.