Revised 2026-09-07

Reading

I list subjects, not titles. Partly because I lend books out and get back a different edition with somebody else’s pencil in the margin, and partly because a recommendation attached to a title gets copied, while a recommendation attached to a subject makes you go and look for the best version of it yourself.

A university textbook on operating-system internals

Finished, twice, eleven years apart.

The chapter on scheduling is better than anything I have found written since, because it bothers to explain the problem before the algorithms: that the scheduler is making a decision with no information about the future, under a deadline measured in microseconds, and that every clever heuristic is a bet on workloads looking like the ones its author had. Everything about memory in it is dated. The scheduling chapter is not.

A reprinted monograph on queueing theory

Read the first third, skipped to chapter nine, keep going back to chapter nine.

Far more mathematics than I can carry, and I have stopped pretending otherwise. The part that earned its place on my desk is the demonstration of what happens to waiting time as utilisation climbs past eighty per cent - not the formula, the plot. I have shown that single page to four different managers and it has ended the argument every time, which no amount of my own graphs ever did.

The optimisation manual for the processor family on my desk

Not read. Consulted, perhaps thirty times.

Eleven hundred pages, and the way to use it is to go in with an exact question and leave immediately. The idle-state tables and the section on what the package can and cannot do while one core is awake are the two places I return to; the first saved me from publishing a wrong explanation of the 133 µs exit cost in one of the notes.

A book on statistics for performance measurement

Abandoned at page 212, and I am not proud of it.

Everything before page 212 was useful enough that I changed how I report numbers: medians of repeated windows instead of averages of one, the sample count stated next to every figure, and the habit of writing down what I think the measurement error is before I take the measurement. Then it turns into a textbook on experiment design for people with more time than me. I will get back to it in a winter with less work in it.

A collected volume of post-incident write-ups

Finished, slowly, one at a time over about two years.

Operators publishing what broke and why, with the organisational parts left in rather than edited out. Two thirds of the failures have nothing to do with the thing that failed; they are about somebody not being told, or a measurement nobody questioned, or a safety mechanism switched off for a good reason in 2019 by a person who had since left. I reread one of these before writing any report, as a reminder that the sentence «and then we assumed» belongs in the report too.

A practical book on electrical installation in buildings

Finished the four chapters that concern me; the rest is for a different trade.

I measure power at the wall for a living and until 2024 I could not have told you honestly what was between my meter and the street. Now I can, which mostly taught me where my own figures stop being trustworthy: the supply’s efficiency curve at low load, and the fact that a «20 W» machine is not drawing twenty watts of anything in particular.

The kernel’s own documentation tree

Not a book. Read like one, in order, over a month.

The power-management and scheduler directories are uneven - some files are better than any article, some have not been true since 2017 - but reading them in order rather than by search gave me the vocabulary to read the source when the documentation runs out. If you only take one thing: the file explaining the idle driver’s governors is short, current, and will change how you read a residency figure.

A six-hundred-page novel about a canal

Eleven months. Finished in March.

Nothing whatsoever to do with any of the above, which is the point. I spent 2024 reading only things that were useful and became measurably worse company.

If you want to argue about any of this, or tell me which edition of the first one you think is best, the contact page is a mail address.