Book guide: Sterbenz & Touch, High-Speed Networking (2001)

The book

James P.G. Sterbenz and Joseph D. Touch (with Julio Escobar, Rajesh Krishnan, Chunming Qiao; technical editor A. Lyman Chapin), High-Speed Networking: A Systematic Approach to High-Bandwidth Low-Latency Communication. Wiley Networking Council Series, 2001. ISBN 0-471-33036-1.
About 500 pages of text, then references (pp. 501–554), Appendix A “Axioms and Principles” (pp. 555–574), Appendix B “Acronyms” (pp. 575–580) and the index.
It is a book of design principles, not a product or protocol manual. It states a small set of axioms and numbered principles in Chapter 2, then restates them for each part of the system: network topology, control, links, switches, hosts, transport, applications.
Its one metric is interapplication delay: the time to move a chunk of data from one application to another, including the time spent inside the applications. Bandwidth matters only because it shortens that delay.
Where the PDF is (Windows machine only): (local cache, not published). The text from pdftotext -layout is next to it as book.txt. The book is copyrighted, so neither file goes into the KB.
Text extraction drops some symbols: µ, ×, +, =, ∞ and the Ø of the first axiom. “104 s” in the text is 104 µs, “D (1 h c)b/r tp” is D = (1 + h + c)·b/r + t_p, and “R )” is R = ∞. When a number in book.txt looks wrong, check the PDF page.

How the principles are numbered

Five axioms: Ø Know the past, present and future; I Application primacy; II High-performance paths goal; III Limiting constraints; IV Systemic optimization. Under IV sit eight numbered design principles (1 Selective optimization … 8 Protocol data units), each with lettered corollaries (1A, 1B, …).

Chapters 3–8 restate the principles for their own part of the system with a letter prefix: N network (Ch 3–4), L link (Ch 5), S switch (Ch 5), E end system (Ch 6), T transport (Ch 7), A application (Ch 8). So E-II.3 is the end-system version of II.3 (store-and-forward avoidance), which becomes “copy minimization”. The full list, in my words, is in 01-principles-map.md.

Chapter map

Ch	Pages	Topic	Read for trading networks?	Notes
1	1–12	Bandwidth vs latency, what “high speed” means, bandwidth–delay product	Skim	02-ch01-02-fundamentals.md
2	13–77	History, the ideal network, constraints, all axioms and design principles, design techniques (cut-through, remapping, pipelining)	Must read: everything else builds on it	02-ch01-02-fundamentals.md
3	79–118	Topology and geography, latency budget, hierarchy, resource trade-offs	Read 3.1 (latency and topology)	03-ch03-topology.md
4	119–164	Signaling and control, connection vs datagram, multicast flows, congestion control, routing dynamics	Read 4.1.6 (multicast) and 4.2	04-ch04-control.md
5	165–284	Links, switches and routers: store-and-forward vs cut-through, buffering, head-of-line blocking, fabrics, multicast in the fabric, lookup, scheduling	Must read 5.2–5.5	05-ch05-links-and-switches.md
6	285–342	End systems: copies, context switches, interrupts vs polling, host–NIC interface	Must read: this is what kernel bypass solves	06-ch06-end-systems.md
7	343–429	End-to-end protocols: connection setup, framing, error control (ARQ vs FEC), flow and congestion control	Read 7.2, 7.4, 7.5	07-ch07-transport.md
8	431–488	Applications: latency classes, masking the speed of light, application–network interaction	Skim	08-ch08-09-applications-future.md
9	489–500	Future directions	Skim	08-ch08-09-applications-future.md

Page ranges checked against the PDF (the PDF page index is the printed page + 20, e.g. printed p. 87 is PDF page 108, index 107).

How to read it in 2026

The principles have aged well. Most examples have not: ATM, SONET, FDDI, active networking, optical burst switching and 2001-era hosts. Read the examples to see a principle at work, not as current practice. The book uses ATM mostly as a warning: Example 2.2 calls the 53-byte cell “an extraordinary example of how not to format and size packets” (p. 70).
The book is written for wide-area networks, where the speed of light dominates. In a colocated trading setup the distances are metres, so the switch hops, the host stack and the application dominate instead. The same principles apply; the terms that matter change (see Selective Optimization, 1A, in the principles map).
Each chapter note has a “Trading-network lens” section. That part is my mapping, not the book. Claims in it are marked: a link if I checked a source, No single source for general engineering knowledge, Unverified otherwise.
The book has a few errors; they are logged in the verification log of the group README (a monitoring-rate figure 10× too low, SONET protection switching printed as 50 µs instead of 50 ms). Principle IDs printed inside chapters sometimes differ from Appendix A; the notes always use the Appendix A IDs.

Ten lessons for trading networks

The book’s principles that matter most for low-latency trading networks, each with the chapter note that explains it. The trading interpretation is my mapping.

Delay is a sum; optimize the biggest term. D = (1 + h + c)·b/r + t_p, and a part that contributes 2% is not worth touching (1A, N-1Al). Measure every hop of tick-to-trade before buying anything. 02, 03
In a colocation, distance is not the problem. Propagation over metres of fiber is nanoseconds, so switch hops, serialization, queuing, copies and the application dominate. Over long routes it is the reverse, and only a straighter or faster medium helps (fiber about 5 µs/km, radio about 3.3 µs/km). 03, 05
Never store and forward, never copy. Cut-through switching pays the serialization time once instead of at every hop; layer-1 switches go further; zero-copy kernel-bypass stacks remove the copies in the host (II.3, S-II.3, E-II.3). 04, 05, 06
A queue is latency, and bursts collide. Keep queues nearly empty and links left of the knee; bursty flows collide at an output port even when the average fits, and correlated bursts defeat statistical multiplexing (N-II.4, Fig. 5.25, Fig. 7.22, p. 407). 05, 07
Packet rate, not bandwidth, sets the design. The time per packet decides what can be done in software and what must be hardware; serial stages must handle the worst case, not the average (S-I.3, E-1Ch, Tables 5.4 and 6.1). 05, 06
Poll when you know data is coming. Interrupts, context switches and user/kernel crossings each cost microseconds; with cheap cores, spinning on the NIC is the right trade (4H, E-II.6c, E-II.6k, 2A). 06
Multicast is a resource trade-off, and its tree is state. One copy per link instead of n; switches replicate in hardware; group joins and trees are state that must be set up before it is needed and kept alive (2C, Example 3.4, N-6B, p. 147). 03, 04, 05
Prefer open loop when round trips are expensive. Redundant A/B paths, snapshots (a datacycle) and FEC recover without waiting; any recovery that needs a round trip costs at least an RTT (6D, T-6De, Fig. 7.6, p. 464). 07, 08
Late is as bad as lost. For real-time data, do not trade timeliness for reliability; a packet far out of order is no better than a lost one (T-I.2, p. 371). 07
Pay setup costs before the open, and re-check trade-offs as prices change. Sessions, joins and connections belong outside the critical path (5A, 6A); the right design moves as bandwidth, processing and memory change in relative cost (Ø4, 2A). 02, 08

Progress

Source: knowledge base note low-latency/00-book-guide.md — own-words notes with sources, projected at build time.